citation: muthu, rajesh (2016) development of a secure...
TRANSCRIPT
Citation: Muthu, Rajesh (2016) Development of a Secure Biometric Recognition System. Doctoral thesis, Northumbria University.
This version was downloaded from Northumbria Research Link: http://nrl.northumbria.ac.uk/31618/
Northumbria University has developed Northumbria Research Link (NRL) to enable users to access the Universityโs research output. Copyright ยฉ and moral rights for items on NRL are retained by the individual author(s) and/or other copyright owners. Single copies of full items can be reproduced, displayed or performed, and given to third parties in any format or medium for personal research or study, educational, or not-for-profit purposes without prior permission or charge, provided the authors, title and full bibliographic details are given, as well as a hyperlink and/or URL to the original metadata page. The content must not be changed in any way. Full items must not be sold commercially in any format or medium without formal permission of the copyright holder. The full policy is available online: http://nrl.northumbria.ac.uk/policies.html
ii
Development of a Secure Biometric
Recognition System
Rajesh Kumar Muthu
A thesis submitted in partial fulfilment
of the requirement of the
University of Northumbria at Newcastle
for the degree of
Doctor of Philosophy
Research undertaken in the Faculty of Engineering and Environment,
Northumbria University, United Kingdom
April 2016
iii
Abstract
Biometric based security systems are becoming an integral part of many security
agencies and organisations. These systems have a number of applications ranging
from national security, law enforcement, the identification of people, particularly for
building access control, the identification of suspects by the police, driverโs licences
and many other spheres. However, the main challenge is to ensure the integrity of
digital content under different intentional and non-intentional distortions; along with
the robustness and security of the digital content.
This thesis focuses on improving the security of fingerprint templates to allow
accurate comparison of the fingerprint content. The current methods to generate
fingerprint templates for comparison purposes mostly rely on using a single feature
extraction technique such as Scale Invariant Feature Transform (SIFT) or Fingerprint
Minutiae. However, the combination of two feature extraction techniques (e.g.,
SIFT-Minutiae) has not been studied in the literature.
This research, therefore, combines the existing feature extraction techniques, SIFT-
Harris: Feature point detection is critical in image hashing in term of robust feature
extraction, SIFT to incorporate the Harris criterion to select most robust feature
points and SIFT-Wavelet: Wavelet based technique is basically used to provide
more security and reliability of image, SIFT feature with efficient wavelet-based
salient points to generate robust SIFT - wavelet feature that provides sufficient
invariance to common image manipulations. The above said feature detector are
known work well on the natural images (e.g., faces, buildings or shapes) and tests
them in the new context of fingerprint images. The results in this thesis demonstrate
that new approach contributes towards the improvement of fingerprint template
security and accurate fingerprint comparisons.
The fingerprint minutiae extraction method is combined individually with the SIFT-
Harris method, SIFT-Wavelet method and the SIFT method, to generate the most
prominent fingerprint features. These features are post-processed into perceptual
hashes using Radial Shape Context Hashing (RSCH) and Angular Shape Context
iv
Hashing (ASCH) methods. The accuracy of fingerprint comparison in each case is
evaluated using the Receiver Operating Characteristic (ROC) curves.
The experimental results demonstrate that for the JPEG lossy compression and
geometric attacks, including rotation and translation, the fingerprint template and
accuracy of fingerprint matching improved when combinations of two different
Feature extraction techniques are used, in contrast to using only a single feature
extraction technique.
The ROC plots illustrates the SIFT-Harris-Minutiae, SIFT-Wavelet-Minutiae, SIFT-
Minutiae perform better than the SIFT method. The ROC plots further demonstrate
that SIFT-Harris-Minutiae outperform all the other techniques. Therefore, SIFT-
Harris-Minutiae technique is more suitable for generating a template to compare the
fingerprint content.
Furthermore, this research focuses on perceptual hashing to improve the minutiae
extraction of fingerprint images, even if the fingerprint image has been distorted. The
extraction of hash is performed after wavelet transform and singular value
decomposition (SVD). The performance evaluation of this approach includes
important metrics, such as the Structural Similarity Index Measure (SSIM) and the
Peak Signal-to-Noise Ratio (PSNR). Experimentally, it has confirmed its robustness
against image processing operations and geometric attacks.
vi
Acknowledgement
The research was carried out in the Faculty of Engineering and Environment at the
University of Northumbria. I would like to express my sincere gratitude to the
university for providing me with a studentship to perform the research. I would also
like to thank VIT University, India for providing sponsorship throughout my
research.
My special thanks go to my supervisors Prof. Ahmed Bouridane and Dr. Fouad
Khelifi for providing me with their continuous support, guidance and
encouragement, and for giving me the opportunity to undertake the study under their
supervision.
I would also like wish to express my sincere gratitude to Prof. Fary Ghassemlooy,
Professor and Head of the Northumbria Communication Research Lab for his
inspiration and support whilst carrying out my research at the University of
Northumbria.
In addition, I would like to convey my sincere thanks to Dr. Richard Binns, Head of
the Department of Physics and Electrical Engineering and to Ms. Karen Vacher,
Office of Postgraduate Research, for their support with my teaching and with the
research.
Finally, special thanks must go to my wife Dr. Rani and my beautiful daughter
Muthusri for their constant support, patience and encouragement throughout this
work.
vii
Publications
1. Muthu, R., Bouridane, A., & Khelifi, F. (2014). Minutiae Based Fingerprint
Image Hashing. International Conference on Control Decision and
Information Technologies, 2014.
2. Herald, S. Muthu, R. Khelifi, F. Ali, A.Vickers, P. & Bouridane, Tanougast,
C. (2013). Progress in Data Encryption Research.
3. Muthu, R., Bouridane, A., & Khelifi, F. (2013). Securing Biometric
Recognition systems. 4th European Workshop on Visual Information
Processing, 2013.
4. Muthu, R., Bouridane, A., & Khelifi, F. (2013). Minutiae Ranking and Its
Application to Fingerprint Recognition. International Journal of Applied
Engineering Research, 8(19), 2013.
5. Muthu, R., Bouridane, A., & Khelifi, F. (2013). Image Feature Extraction
Techniques for Biometric Recognition Systems-A Survey. International
Journal of Applied Engineering Research, 8(19), 2013.
6. Muthu, R., Bouridane, A., & Khelifi, F. (2013). Minutiae Ranking and Its
Application to Fingerprint Recognition. International Conference on
Communications, Networking and Signal Processing, 2013.
7. Muthu, R., Bouridane, A., & Khelifi, F. (2013). Image Feature Extraction
Techniques for Biometric Recognition Systems -A Survey. International
Conference on Communications, Networking and Signal Processing, 2013.
viii
Declaration
I declare that the work contained in this thesis has not been submitted for any other
award and it is my own work.
Rajesh Kumar Muthu
ix
Contents
Abstractโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ iii
Acknowledgementโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ... vi
Publicationโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ... vii
Declarationโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ... viii
List of figuresโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ... xiii
List of tablesโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. xvi
List of acronymsโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ...... xvii
1 Introductionโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. 1
1.1 Biometric Recognition Systemโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. 1
1.1.1 System Operationโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.. 2
1.2 System Vulnerabilityโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. 4
1.3 Fingerprint Representationโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ 6
1.3.1 Fingerprint Minutiae Extraction ........................................................... 9
1.4 Perceptual Fingerprint Image Hashing ................................................. 9
1.4.1 Properties of Perceptual Image Hashingโฆ.. ........................................ 12
1.5 Aim and Objectives............................................................................... 15
1.6 Thesis Contributionโฆโฆ....................................................................... 16
1.7 Organization of the thesis...................................................................... 16
2 Review of Secure Biometric Systemโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.. 18
2.1 Minutiae Based Fingerprint Systemโฆ.. ............................................... 18
2.2 Feature Extraction and Image Hashing Techniquesโฆโฆโฆโฆโฆโฆโฆ.. 20
2.3 Temple Protection Techniques.............................................................. 22
x
2.4 Summary............................................................................................... 26
3 Image Representation In The Transform Domainโฆ....................... 27
3.1 Discrete Fourier Transformโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ... 28
3.2 Fourier-Mellin Transformโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. 30
3.3 Discrete Cosine Transform โฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ... 31
3.4 Discrete Wavelet Transformโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. 32
3.4.1 Multiresolution Analysisโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.. 33
3.4.2 Propertiesโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ... 38
3.4.3 2-D wavelet transformโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ 39
3.5 Summaryโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ............................... 40
4 Image Feature Extraction Techniques............................................... 42
4.1 Introduction...................................................................................... 42
4.2 End-Stopped Wavelets.......................................................................... 43
4.2.1 Experimental Results.............................................................................. 46
4.3 Speed Up Robust Feature....................................................................... 48
4.3.1 Experimental Results............................................................................. 50
4.4 Scale Invariant Feature Transform...................................... โฆโฆโฆ 54
4.5 SIFT-Harris.............................................................................. โฆโฆโฆ 55
4.5.1 Experimental Results........................................................................... 56
4.6 SIFT Wavelet............................................................................... 58
4.7 Overview of feature Detector................................................... โฆโฆโฆ 58
4.8 Summary............................................................................................... 62
5 Minutiae Based Fingerprint Image Hashing..................................... 63
5.1 Design criteria for image hashing......................................................... 64
xi
5.2 Perceptual Image Hashing Schemes..................................................... 65
5.2.1 Perceptual Image Hashing Framework................................................ 66
5.3 Content-Based Image Authentication................................................... 67
5.4 Proposed Feature Extraction Techniques............................................. 68
5.4.1 SIFT -Harris-Minutiae Feature for Fingerprint Image Hashing........... 69
5.4.2 SIFT-Wavelet -Minutiae Feature for Fingerprint Image Hashingโฆโฆ 70
5.4.3 SIFT-Minutiae Feature for Fingerprint Image Hashing......................... 74
5.5 Image Hashing Based on Shape Contexts.......................................... .. 75
5.5.1 Shape Contexts...................................................................................... 75
5.5.2 Shape Context Based Hashing............................................................... 76
5.6 Experimental Results.............................................................................. 80
5.7 Summary................................................................................................ 97
6 Perceptual Hashing To Improve Minutiae Extraction of
Fingerprint Imagesโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.
98
6.1 Singular Value Decomposition (SVD)โฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. 101
6.1.1 Properties of SVDโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. 102
6.2. Proposed Hash Securing For Fingerprint Images.โฆโฆโฆโฆ.โฆโฆโฆ... 102
6.3 Performance Evaluation of Proposed Approachโฆโฆโฆโฆ... ... โฆโฆ. 104
6.3.1 Structural Similarity Index Measure (SSIM)โฆโฆโฆโฆโฆโฆ.. โฆโฆโฆ. 104
6.3.2 Peak Signal-to-Noise Ratio (PSNR)โฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ..... 105
6.3.3 Bit- Error- Rate (BER)โฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ โฆโฆโฆโฆ. 105
6.4 Experimental Resultsโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ 106
6.5 Summaryโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.. โฆโฆโฆโฆโฆโฆโฆโฆโฆ 112
7 Conclusions and future work........................................................... 113
7.1 Introduction ........................................................................................ 113
xii
7.2 Contribution of the thesis...................................................................... 113
7.3 Future work........................................................................................... 115
Bibliographyโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ 118
xiii
List of Figures
Figure 1.1 Example of some of commonly used biometric traitsโฆโฆโฆโฆโฆ. 2
Figure 1.2 Enrollment and Authentication stage of a biometric systemโฆโฆ... 4
Figure 1.3 Vulnerabilities in a biometric system [6]....................................... 6
Figure 1.4 A fingerprint image with core and minutiae points marked on it.
The global structure is the ridge pattern along with the core and delta points
and local structure are characterized by minutiae pointsโฆโฆโฆโฆโฆโฆโฆโฆ
7
Figure 1.5 Minutiae of fingerprint templateโฆโฆโฆโฆโฆ............................... 7
Figure 1.6 Example of fingerprint system application: (a) Border passage
system using fingerprint system. (b) Fingerprint Identification system
in ATM machine. (c)Walt Disney world use fingerprint recognition system
for annual and seasonal pass holders to access into the park. (d) Fingerprint
โ based point of sale. (e)Fingerprint -based door lock. (f) Fingerprint
system is used in time and attendance applications. (g) Fingerprint
verification in mobile phoneโฆโฆโฆโฆโฆ..โฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ...
8
Figure 1.7 Minutiae extraction of the fingerprint image: (a) Fingerprint
image. (b)Thinned image. (c) Minutiae extraction. (d) Minutiae and its
spurious. (e) Removal of false minutiae. (f) Final Minutiae of fingerprint
imageโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.
9
Figure 1.8 Perceptual hashing for fingerprint image authentication โฆโฆโฆโฆ 10
Figure1.9 Architecture for three way check for template Protection through
Perceptual hashingโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.
11
Figure 1.10 Examples of fingerprint image under different content
preserving operationsโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.โฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ...
14
Figure 1.11 Examples of perceptually different images................................. 15
Figure 2.1 Template protection Schemes [5].................................................. 23
Figure 3.1 One-stage wavelet decompositionโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ... 36
Figure 3.2 Two-stage wavelet decompositionโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.. 36
Figure 3.3 One-stage wavelet reconstructionโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ 38
xiv
Figure 3.4 Two-stage wavelet reconstructionโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ... 38
Figure 3.5 One-stage 2-D wavelet decompositionโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.. 40
Figure 3.6 One-stage 2-D wavelet decomposition of โLenaโโฆโฆโฆโฆโฆโฆ.. 40
Figure 4.1 Behaviour of the end-stopped wavelet on a synthetic image. (a)
Synthetic L-shaped image. (b) Response of a Morlet wavelet. (c) Response
of the end-stopped wavelet....................................................................
45
Figure 4.2 Sample of fingerprint images from FVC2004/DB1_A database.... 47
Figure 4.3 Approximation of the second order Gaussian partial derivative.
The grey regions are equal to zero...................................................................
49
Figure 4.4 Feature Descriptors and Matching of Feature Points (30Points)
on Fingerprint Images of the same subject for different attacks: (a) Feature
Descriptor (b) Feature Points (379 Points) (c) Feature Points (d) 5 degree
Rotation e) 20 degree Rotation (f) 180 degree Rotation (g) Translate
(25x25) (h) Histeq (i) Median Filter (5x5) (j)JPEG (10%) (k) AWGN
(0.065%)..........................................................................................................
52
Figure 4.5 Feature Descriptors and Matching of Feature Points (30 Points)
on Lena Images of the same subject for different attacks: (a) Feature
Descriptor (b) Feature Points (379 Points) (c) Feature Points (d) 5 degree
Rotation e) 20 degree Rotation (f) 180 degree Rotation (g) Translate
(25x25) (h) Histeq (i) Median Filter (5x5) j) JPEG (10%) (k)AWGN
(0.065%)...........................................................................................................
53
Figure 5.1 Design requirements for fingerprint image hashingโฆโฆโฆโฆโฆโฆ 65
Figure 5.2 Pipeline stages of a perceptual hashing system............................ 67
Figure 5.3 Proposed Robust SIFT-Harris- Minutiae based fingerprint image
hashing.............................................................................................................
69
Figure 5.4 Second Level Wavelet Transform................................................... 73
Figure 5.5 Image Decompose Levelโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. 73
Figure 5.6 Proposed Robust SIFT-Wavelet- Minutiae Fingerprint Image
Hashingโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.................................................................
74
Figure 5.7 Diagram of the original shape contexts and the proposed shape
contexts hashing: RSCH.ASCH. (a) Original Shape Contexts. (b)Radial
76
xv
Shape Contexts hashing (c) Angular shape Contexts hashing.......................
Figure 5.8 Radon transform ๐ (p,๐) of a 2-D function ๐(๐ฅ,๐ฆ)......................... 79
Figure 5.9 Fingerprint images from FVC2002/DB1_A database..................... 82
Figure 5.10 ROC curves of the proposed robust minutiae of fingerprint
image (SIFT-Harris-Minutiae) using shape context based image hashing
technique, (a) ROC curves under JPEG lossy compression (b) ROC curves
under median filter (c) ROC curves under Gaussian blur (d) ROC curves
under rotation (e) ROC curves under
translation................................................................................................
87
Figure 5.11 ROC curves of the proposed robust minutiae of fingerprint
image(SIFT-Wavelet-Minutiae) using shape context based image hashing
technique, (a) ROC curves under JPEG lossy compression
(b) ROC curves under median filter (c) ROC curves under Gaussian
blur (d) ROC curves under rotation (e) ROC curves under
translationโฆโฆ............................................................................................
90
Figure 5.12 ROC curves of the proposed robust minutiae of fingerprint
image(SIFT-Minutiae) using shape context based image hashing technique,
(a) ROC curves under JPEG lossy compression (b) ROC curves under
median filter (c) ROC curves under Gaussian blur (d) ROC curves under
rotation (e) ROC curves under Translationโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ..
93
Figure 5.13 ROC curves of the proposed robust minutiae of fingerprint
image(SIFT) using shape context based image hashing technique, (a) ROC
curves under JPEG lossy compression (b) ROC curves under
median filter (c) ROC curves under Gaussian blur (d) ROC curves under
rotation (e)ROC curves under translation.......................................................
96
Figure 6.1 Proposed hash securing for fingerprint imagesโฆโฆโฆโฆโฆโฆโฆโฆ 103
Figure 6.2 Average SSIMโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ. 111
Figure 6.3 Average PSNRโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ.... 111
Figure 6.4 Average Hausdorff Distances of Minutiae...................................... 111
xvi
List of Tables
Table 2.1 Different methods used to transform fingerprint features for
template protection [63].................................................................................
25
Table 4.1 Different Attacks used to assess the End Stopped Feature
pointโฆโฆ
47
Table 4.2 Hausdorff Distance between features of original and Attacked
Image (for 32 Feature Points)โฆโฆโฆโฆโฆ.......................................................
48
Table 4.3 Different Attacks used to assess the SURF Feature
point..................................................................................................................
51
Table 4.4 Comparison of keypoints Detectors on original and distorted
imageโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ
57
Table 4.5 Average Hausdorff Distance between the coordinate for the top 20
keypoints detected in the original and manipulated copies using the SIFT-
harris and End Stopped Detector..................................................................... .
57
Table 4.6 Overview of Feature Detector........................................................... 61
Table 5.1 Content-Preserving and Content-Changing Manipulationโฆโฆ........ 68
Table 5.2 Different Attacks used to assess the hashing performanceโฆโฆ....... 81
Table 5.3 Summary of proposed fingerprint image hashing technique........ 84
Table 6.1 Different content persevering operation used to assess the hash
Performanceโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆโฆ..
108
Table 6.2 Average SSIM and PSNR Valueโฆโฆโฆโฆโฆโฆโฆโฆ..................... 109
Table 6.3 Average Hausdorff Distance of Minutiaeโฆโฆโฆโฆโฆโฆโฆ............. 110
Table 6.4 Performance Evaluation of Metrics and Minutiaeโฆ....................... 110
xvii
List of Acronyms
AFIS : Automatic Fingerprint Identification Systems
AWGN: Additive White Gaussian Noise
ASCH : Angular Shape Context Hashing
CL : Constract Low
CH : Constract High
DCT : Discrete Cosine Transform
DFT : Discrete Cosine Transform
DWT : Discrete Wavelet Transform
DoG : Difference-of-Gaussian
EER : Equal Error Rate
FDoG : First Derivative of the Gaussian
FMT : Fourier- Mellin Transform
FPR : False Positive Rate
FP : Feature Points
HVS : Human Visual System
LSH : Locality-Sensitive Hashing
MCC : Minutiae Cylinder Code
PSNR : Peak Signal-to-Noise Ratio
RSCH : Radial Shape Context Hashing
ROC : Receiver Operating Characteristic
RLRD : Random Local Region Descriptor
xviii
SIFT : Scale Invariant Feature Transform
SVD : Singular Value Decomposition
SSIM : Structural Similarity Index Measure
SURF : Speeded Up Robust Feature
SNR : Signal-Noise-Ratio
TPR : True Positive Rate
1
Chapter 1
Introduction
1.1 Biometric Recognition System
Identification of a personal identity in a digital environment can be established in
three basic ways i.e. by โsomething you knowโ (e.g. password, PIN number etc), by
โsomething you carryโ (e.g. ID cards, keys etc) or โsomething you areโ (e.g.
fingerprints, face, iris etc). The security and recognition systems based on surrogate
representations, for instance passwords and ID cards have been established to have a
fundamental flaw, for example a password can be forgotten or guessed, an ID card
can easily be lost or misplaced and they can all be very easily spoofed.
Biometric based recognition systems are being extensively used in a number of
current and potential applications ranging from national security, law enforcement,
the identification of people, particularly for building access control, the identification
of suspects by the police, driverโs licences and many other spheres. Therefore,
current trends in the development of innovative security systems, particularly
pertaining to the identification and verification of an individual, places considerable
emphasis on biometric based solutions for the reason that with biometric
identification systems the key is the user and in most cases very difficult to forget.
Biometric technologies can be defined as โautomated methods for verifying or
recognizing the identity of a person based on a physiological and /or behavioural
characteristicโ [1]. Figure 1.1 signifies commonly used biometric traits including
fingerprints, face, iris, palm print, signature and voice [2],[3].
2
(a) Fingerprint (b) Face (c) Iris
(d) Palmprint (e) Signature (f) Voice
Figure1.1: Example of some commonly used biometric traits
1.1.1 System Operation
A biometric system, regardless of the algorithms, consists of four major modules:
Sensor Module, Feature Extraction, Matching Module and Decision Module [4].
The Sensor Module captures the biometric data of a user. For example, a
fingerprint sensor that captures the fingerprint impression of a user.
The Feature Extraction Module, where the captured data is processed to
extract feature sets. For example, the position and orientation of minutiae in a
fingerprint image would be computed in the feature extraction module of a
fingerprint system.
The Matching Module compares the feature sets against those in the system
database by generating a matching score. For instance, the number of
matching minutiae between the query and the template can be computed as a
matching score.
3
The Decision Module, where the userโs claimed identity is either a
match or non-match, if the match score is greater than the system
threshold and if not declares a non-match.
A biometric authentication system is essentially a pattern recognition system that
recognizes a person by determining the authenticity of biometric traits. A biometric
system operates in three main stages: (i) Enrollment: the system collects biometric
data from a user and features extracted from the data is stored as a biometric
template, (ii) Verification: the system authenticates a personโs identity by comparing
the captured biometric data with previously enrolled biometric reference template
pre-stored in the system. It conducts one-to-one comparison to confirm whether the
claim of identity by the individual is true. (iii) Identification: the system recognizes
an individual by searching the entire enrollment template database for a match. It
conducts one-to-many comparisons to establish if the individual is present in the
database and if so, returns the identifier of the of the enrollment reference that
matched [5][6]. Figure 1.2 illustrates the enrollment, verification, and identification
stage of a biometric recognition system
Biometric
Sensor
Quality
Checker
Feature
Extraction
System
Database
User Identity
a) Enrollment
User
Biometric
Sensor
Quality
Checker
Feature
Extraction
System
Database
Claimed Identity
b) Verification
Matcher
Decision
Module
Match/Non
Match
User
4
Biometric
Sensor
Quality
Checker
Feature
Extraction
System
Database
c) Identification
Matcher
Decision
Module
Match/Non
Match
User
Figure 1.2: (a) Enrollment, (b) verification, and (c) Identification stage of a
biometric recognition system.
1.2 System Vulnerabilities
A biometric system is vulnerable to different types of attacks that can compromise
the systemโs security. The various factors that affect the security of the system
typically belong to one of four categories: intrinsic failures, administrative attacks,
non-secure infrastructure and access to biometric data [7], [8].
Intrinsic failures: It is a security lapse due to an incorrect decision made by the
biometric system. The sensor may fail to acquire the biometric data of the user
due to limits in sensing technology or environmental conditions. The variation in
the imaging condition captured biometric data and thus, the features extracted
usually exhibit considerable inter-user similarity and intra-user variations. e.g., the
face images of two identical twins are very similar to each other and this may lead
to an incorrect decision when verifying the identity of one of the twins. The error
rate at which the biometric verification system incorrectly matches two unrelated
biometric templates is called the โfalse accept rateโ. Conversely, it may also fail to
match two biometric templates extracted from the same biometric due to
significant intra-user variations. These kinds of errors are measured using the
โfalse reject rateโ system.
Administration attacks: This refers to all vulnerabilities due to improper
administration of the biometric system. The function of the system can be
5
abused by an attacker by colluding with or coercing a system administrator to
allow the individual to enrol or be accepted as a genuine user.
Nonโsecure infrastructure: where an adversary can manipulate the biometric
infrastructure in the hardware, software and the communication channels
between the various modules.
Access to biometric traits: An adversary covertly captures the biometric data of
the legitimate user and uses the data to create physical artefacts. Hence, if the
system is not capable of distinguishing between a live biometric and an artificial
spoof, an adversary can circumvent the system by presenting spoofed traits.
As such biometric systems are prone to vulnerability at different points in the system
[9] (Figure 1.3). These attacks are intended to either circumvent the security
provided by the system or to change the normal functioning of the system:
(i) A fake biometric trait such as an artificial finger may be presented at the
sensor
(ii) Illegally intercepted data may be resubmitted to the system.
(iii) The feature extractor may be replaced by a Trojan horse program that
produces predetermined feature sets.
(iv) Legitimate feature sets may be replaced with synthetic feature sets.
(v) The matcher may be replaced by a Trojan horse program that always
outputs high scores, thereby defying the system security
(vi) The template stored in the database may be modified or removed.
Alternately, a new template may be introduced in the database.
(vii) The data in the communication channel between various modules in the
system may be altered.
(viii) The final decision output by the biometric system may be overridden.
6
SensorFeature
ExtractorMatcher
Application device
(e.g., cash dispenser)
Stored
Template
1. Fake Biometric
2. Replay Old
Data
3. Override
Feature Extractor
4. Synthesized
Feature Vector5. Override Matcher
6. Modified template
8. Override Final Decision
7. Intercept
the Channel
Yes/No
Figure1.3: Vulnerabilities in a biometric recognition system [8].
1.3 Fingerprint Representation
Amongst all biometric traits, fingerprints are the oldest serving, most successful and
popular modality to identify a person. Fingerprints consist of a regular texture pattern
composed of ridges and valleys. In a fingerprint the focussed feature points are the
minutiae. i.e., ridge endings and ridge bifurcation [1] (Figure 1.4). The spatial
distribution of these minutiae points is said to be unique for each finger and thus, the
collection of minutiae points in a fingerprint is primarily employed for matching two
fingerprints. A good quality fingerprint consists of between 20 to 70 minutiae [10],
all of which are not genuine and prominent. The minutiae of the fingerprint image
are usually represented as a 3 โtuple (๐ฅ, ๐ฆ, ๐), where ๐ฅ and ๐ฆ are the coordinate of
the minutiae and ๐ is the angle (Figure 1.5).
7
Figure 1.4: A fingerprint image with core and minutiae points marked on it. The
global structure is the ridge pattern along with the core and delta points. Local
structures are characterized by minutiae points [4].
Figure 1.5: Minutiae of fingerprint template
X y ฮธ
114 135 90
90 113 80
167 161 110
128 66 120
Ridge ending
Ridge bifurcation
Minutiae of Fingerprint Image Fingerprint Template
(Minutiae)
()
Ridge ending
Ridge bifurcation
Core
8
(a) (b) (c)
(d) (e)
(f) (g)
Figure 1.6: Example of fingerprint system application: (a) Border passage
system using the fingerprint system. (b) Fingerprint identification system in an
ATM machine. (c)Walt Disney World use the fingerprint recognition system for
annual and seasonal pass holders to access the park. (d) Fingerprint โ based point
of sale. (e) Fingerprint - based door lock. (f) The fingerprint system is used in
time and attendance applications (g) Fingerprint verification on a mobile phone
[4].
9
1.3.1 Fingerprint Minutiae Extraction
Most Automatic Fingerprint Identification Systems (AFIS) (Figure1.6) are based on
minutiae matching. A minutiae based fingerprint recognition system undergoes three
main stages, namely pre-processing, minutiae extraction and post-processing [1], [2].
The pre-processing stage consists of image enhancement, image binarization and
image segmentation. Thinning and minutiae marking is completed at the minutiae
extraction stage, and finally, the removal of false minutiae in the post-processing
stage. Figure 1.7 illustrates minutiae extraction of the fingerprint image.
(a) (b) (c)
(d) (e)(f)
Figure 1.7: Minutiae extraction of the fingerprint image: (a) Fingerprint image. (b)
Thinned image. (c) Minutiae extraction. (d) Minutiae and its spurious. (e) Removal
of false minutiae. (f) Final minutiae of fingerprint image.
1.4 Perceptual fingerprint image hashing
Biometric template security is an important and emerging research, given that unlike
passwords and tokens, compromised biometric templates cannot be revoked and
reissued. The main challenge is to ensure the integrity of digital content under
different intentional and non-intentional distortions environment together with the
robustness and security of the data content [7], [8], [11], [12]. Recently, a few
methods have been proposed for biometric template security. These methods can be
categorized into two classes: Transformed based systems and biometric
10
cryptosystems [7]. These approaches offer a positive solution to the problem of user
authentication, although they are limited. In addition, existing approaches are also
not robust enough for geometric operations and slight variations in the template can
significantly decrease the performance.
Alternatively, image hashing, a scheme that generates a unique, compact, resilient
and secure signature for each image, has been widely applied in image content
verification and content authentication.
Pre-
Processing
Feature
Extraction
Post -
Processing
Image
Hashes
ComparisonQuery Hashes
Secret Key
Database
Verification /Identification
Fingerprint Image
sets
Figure 1.8: Perceptual hashing for fingerprint image authentication
A perceptual hashing system for fingerprint image authentication, as illustrated in
Figure 1.8, generally consists of three main stages: pre-processing, feature extraction
and post-processing. An image hash can be constructed by extraction and post
processing appropriate image features to form a compact representation that can be
used for the authentication and integrity of the data [13]. Interestingly another
advantage of a hash based image authentication scheme is that it can also be used to
handle key issues like tamper detection, security and robustness. The robustness of
an image hashing arises from robust feature extraction and the compression, which
mainly contributes to the compactness of the final hash. To increase the security of a
traditional hash function and prevent unauthorized access, a secret key is
incorporated in the feature extraction, the compression or both to make the hashes
unpredictable.
11
Most of hashing algorithms incorporate a pseudorandomization relying on a secret
key into the compression step [14][15][16] to further enhance the security, as
indicated by the dashed line in Figure 1.8. The key is owned by the owner, and the
hash generation is a pseudorandom process rather than a completely random one for
fingerprint identification. The incoming query hash corresponding to the query image
is compared with the hashes in the database.
The approach to perceptual hashing is extended to apply and demonstrate the three
way check [17] to protect the biometric template as shown in Figure 1.9.
Smart Card Hi(hs)==Hb(hs)
Secured Database
Hs(h1)
Hs(h2)
Hs(h2)
โโ
โโ
โโ
Hs(hn)
Kiosk
FP, Hi(hs) Claim Hb(hs)?
yes
Biometric
Measurement
FP
Hb(hs)
Feature Extraction
Genarate Hash
Input image
Figure1.9: Architecture for the three way check for the template Protection through
Perceptual hashing.
The architecture for the three way check for the template protection through
perceptual hashing explains the biometric template in the form of a selected feature
point (FP). Furthermore, the hash template FP, Hi hs is stored in the smartcard/e-
passport and the reference hash template๐ป(๐๐ ) is stored in a secured database. The
12
three way check is performed by matching the hashed template form, the smartcard
database and biometric measurements i.e. Hi hs = ๐ป(๐๐ ) = ๐ป๐(๐๐ )
The three way check proceeds as follows:
The Kiosk, which is the border control authority that reads the information
feature points (only selected/limited points) and the hashed template from the
smart card/e-passport sends the feature points to the biometric measurements.
This results in the output of the feature point combining with the FP to
generate the ๐ป๐ (๐๐ ) hash template, which is given to the Kiosk.
The Kiosk then validates the received hash template of Hi hs and the
biometric measurement hashed template ๐ป๐(๐๐ ) with the database hash
template ๐ป(๐๐ ), to check the authenticity of the owner i.e.
Authenticity: Hi hs = ๐ป(๐๐ ) = ๐ป๐ (๐๐ ) otherwise it is not authentic.
1.4.1 Properties of perceptual image hashing.
Given an image I and its perceptually similar copy with minor distortion ๐ผ๐ . Let ๐, ๐
be two positive values that satisfy ๐ > 0, ๐ < 1. The image hashing function ๐ป๐ (.)
depends on the secret key k. The desirable properties of a perceptual fingerprint
image hashing function ๐ป๐ (.) are as follows:
One-way function: Preferably, the hash generation should be noninvertible
I โ ๐ป๐ (.) (1.1)
Compactness: The size of the hash signature ๐ป๐ (๐ผ) should be much smaller
than that of the original image I.
๐๐๐ง๐ ๐ป๐ (๐ผ) โช ๐๐๐ง๐ (๐ผ) (1.2)
Perceptual Robustness: The robustness property requires for any pair of
perceptually similar images have similar hashes even if they undergo content-
preserving operations. This is for image identification proposes. An example
is illustrated in Figure 1.10 (a) to (i), which includes the original image and
its distorted copies under distortions such as rotation, median filter, Gaussian
blur, Gaussian noise, JPEG compression, motion blur, translation and average
13
filter. The perceptual robustness of image hashing guarantees that these
images will have very similar hashes.
๐ป๐ (๐ผ) โ ๐ป๐ ๐ผ๐ โฅ 1 โ ๐, 0 โค ๐ < 1. (1.3)
Visual Fragility: Perceptually distinct images (Figure 1.11) should have
different hashes.
๐ป๐ (๐ผ) โ ๐ป๐ ๐ผโฒ โฅ 1 โ ๐, 0 โค ๐ < 1. (1.4)
Unpredictability:
๐ป๐ (๐ผ) ; ๐๐ (1) โ ๐๐(0) โ 0.5 (1.5)
Where ๐๐(๐ฅ) is the probability mass function for ๐. With this property the hash
values should be approximately equally distributed. Security is an important concern
for image hashing. Pseudo-randomisation techniques are incorporated into the image
hash generation process to enhance the security of image hashes by using secrete
keys.
14
(a) Original fingerprint Image (b) Rotation (c) Median filter
(d) Gaussians blur (e) Gaussian Noise f) JPEG Lossy compression
(g) Motion Blur (h) Translation (i) Average Filter
Figure 1.10: Examples of fingerprint images under different content preserving
operations (perceptually insignificant modifications).
15
Figure 1.11: Examples of perceptually different images
1.5 Aim and Objectives
The aim of this research is to investigate and develop novel perceptual hashing
techniques to secure biometric templates whilst maintaining the fingerprint
recognition rate. In addition, a technique to improve minutiae extraction using
perceptual hashing in fingerprint identification systems is also studied. The main
objectives are as follows:
To analyse the performance of state of the art techniques, SIFT, SIFT-Harris
and to improve the feature extraction of fingerprint images.
To develop a technique for fingerprint image hashing to improve the security
of biometric templates, as well as to enhance or maintain the performance of
the recognition system. Furthermore, the trade-off between robustness and the
security of the hash will be addressed.
To develop a method for minutiae extraction of fingerprint images using
perceptual hashing that tolerates content-preserving manipulation.
Furthermore, evaluate the imperceptibility against the content preserving
operations using standard metrics such as SSIM and PSNR.
16
1.6 Thesis Contribution
The fingerprint image hashing and the related security issues are improved in this
thesis by focusing on robust feature extraction techniques. The following are the
three important contributions of this thesis:
A new, robust, SIFT-Harris-Minutiae feature extraction technique that
includes orientation and descriptor in the minutiae of fingerprint images using
SIFT-Harris feature points for improving robustness against image processing
operations including JPEG lossy compression and geometric attacks such as
rotation and translation.
A new, robust, SIFT-Wavelet-Minutiae feature extraction technique for
improving robustness to the median filter, Gaussian blurs and rotation
attacks. The ROC plots demonstrate that this new technique is suitable for
generating a secured fingerprint template.
A new robust perceptual hashing solution, based on wavelet transform and
singular value decomposition (SVD) to enhance the fingerprint image and to
provide good balance of robustness and imperceptibility. Additionally, this
approach retains the maximum minutiae of the fingerprint image, even if the
image is distorted.
1.7 Organisation of the thesis
This thesis comprises seven chapters, which are outlined below:
Chapter 1 presents a brief introduction on perceptual fingerprint image hashing,
including the original contribution of the research.
Chapter 2 discusses the literature review in relation to biometric systems, which
includes the minutiae based fingerprint system, the feature extraction and image
hashing scheme, and template protection techniques.
Chapter 3 discusses the basic characteristics of transform domain: Discrete Fourier
Transform, Discrete Cosine Transform and Fourier-Mellin Transform and Discrete
Wavelet Transform. The transform domain technique are the most challenging part
17
of the image hashing in the feature extraction stage, the extracted feature are
invariant to image processing operation and geometric attacks.
Chapter 4 highlights the state of the art feature extraction processes, which include
the end-stopped wavelet, and the SIFT, SURF and SIFT-Harris methods.
Chapter 5 presents the proposed framework for minutiae based fingerprint image
hashing. The method is combined individually with the SIFT-Harris, SIFT-Wavelet,
SIFT. The accuracy of fingerprint comparison in each case is evaluated using the
ROC curves. Further, the ROC plots demonstrate accuracy of fingerprint matching
improved when combinations of two different feature extraction techniques are used,
in contrast to using only a single feature extraction technique.
Chapter 6 presents the proposed technique to improve minutiae extraction of
fingerprint images using perceptual hashing. The imperceptibly and performance of
the minutiae extracted are discussed in detail.
Chapter 7 discusses the conclusion and suggests areas for future research in
biometric recognition systems.
18
Chapter 2
Review of Secure Biometric Systems
Biometric recognition is often considered to enhance identity verification. The use of
biometric recognition also introduces new challenges in protecting the privacy of the
subject and increases the security of the verification system. The literature review
discusses three major categories: the minutiae based fingerprint system, the feature
extraction and image hashing scheme, and template protection techniques.
2.1 Minutiae Based Fingerprint System
Fingerprint is a popular biometric modality, which is used extensively in several
applications for person recognition, providing uniqueness and an acceptable
performance. A minutiae based fingerprint system involves three basic steps: pre-
processing, feature extraction and matching. This system requires storing minutiae
sets in the database. However, several projects have established that the fingerprint
impression can be reconstructed from minutiae information.
More recently, several researchers have addressed the concept of fingerprint template
solution. Moujahdietat [19] proposed a new approach to fingerprint template by
constructing a new representation of minutiae based on spiral curves. Liang et al [20]
demonstrated a robust fingerprint indexing scheme using a minutiae neighborhood
structure and low order Delaunay triangles. This algorithm was able to search a
fingerprint database more efficiently and methodically for various fingerprints.
Jin et al [21] proposed a fingerprint template protection method that transforms a set
of minutiae points into bit-string using the polar grid based 3 -tuple quantization
technique that offers a reasonable recognition rate. Moreover, Liu et al [22] utilised a
random local region descriptor (RLRD) to generate a fixed-length feature vector. In
this case, the RLRD features are extracted from a set of randomly and uniformly
generated directional points from the image of a fingerprint. Additionally, Feng et al
[23] suggested two descriptors: a texture based descriptor, which captures the
orientation and frequency information of minutiae and a minutiae-based descriptor
matching algorithm. The combined descriptors provide high discriminating ability.
19
Cappelli et al [24] described a template privacy protection technique for minutiae
cylinder code (MCC), which provides diversity, revocability and irreversibility for
the MCC descriptors with respect to the original minutiae, with the aim of improving
recognition accuracy while reducing the size of the template. Tulyakov et al [25] [26]
presented a method of symmetric hashing of the fingerprint minutiae, aimed at
protecting the original fingerprint and minutiae location from the attacker; whereas
Sutcu et al [27] proposed a scheme which employs a robust oneโway transformation
that maps the geometrical configuration of the minutiae points into a fixed-length
code vector. Moreover, Shuai et al [28] recommended locality-sensitive hashing
(LSH) based fingerprint indexing using reduced SIFT points.
Xu et al [29] [30] approached the spectral minutiae representation as a fixed-length
feature vector to represent the minutiae set. In addition, Feng et al [31] presented a
novel fingerprint-matching algorithm that matches both the minutiae and the ridges.
This approach is used to find promising initial minutiae pairs. For each initial
minutiae pair, a ridge matching process was performed, which incrementally
matched the remaining minutiae and ridges.
Additionally, Jain et al [32] described a hierarchical matching system that utilized
features at all three levels, namely, level 1 (pattern), level 2 (minutiae points) and
level 3 (pores and ridge contours). A relative reduction of 20% is observed in the
Equal Error Rate (EER: The rate at which both acceptance and rejection errors are
equal. In general, the device with the lowest EER is the most accurate) of the
matching system, when level 3 features are employed in combination with level 1
and level 2 features. Using a different method, Ahn et al [33] proposed an interesting
alignment-free feature transformation approach. The purpose of this technique is to
extract some special geometrical information from minutiae triplets to construct the
secure template.
Lee et al [34] suggested a method for generating cancellable fingerprint templates
without alignment, in addition to a method for producing changing functions. To
generate the templates for each minutiae, a rotational and translation invariant value
is computed from the orientation information of the neighbouring local region
surrounding the minutiae. The invariant value is used as the input for two changing
20
functions that output two values for the translational and rotational movements of the
original minutiae, respectively, in the cancellable template.
Chang et al [35] proposed a point pattern matching to solve the problem of optimal
matches between a two-point pattern under geometrical transformation and spurious
points pattern. To increase the reliability and robustness of minutiae matching, Jiang
et al [36] recommended a fingerprint minutiae matching by using both the local and
global structures of the minutiae. However, the system determines the identity of a
user by comparing the match score to a threshold value set by the administrator.
Luo et al [37] introduced ridge information into the process of fingerprint matching
and used a changeable sized box in the matching process; whilst Bhowmick et al [38]
presented a method to assign a score value to each of the extracted minutiae, based
on several topographical properties. The score associated to the minutiae signifies its
genuineness and prominence. Furthermore, Jain et al [39] advocate the design and
implementation of an on-line fingerprint verification system, which operates in two
stages, such as minutiae extraction and minutiae matching. To find the
correspondence between minutiae in the input image and the stored template, an
alignment based elastic matching algorithm is employed.
2.2 Feature Extraction and Image hashing Techniques
In the design requirements for fingerprint image hashing, the robustness of the image
hashing arises from robust feature extraction and compression, which mainly
contributes to the compactness of the final hash. The hash should be capable of
dealing with various image processing and geometric attacks, as long as two similar
images possibly generate near similar hash value, otherwise the hash value differs
with a perceptually different image. The following literature discusses the various
feature extractions and image hashing schemes.
Monga et al [40] established a โnonnegative matrix factorizationโ, which is robust
against a significant class of attacks; including blurring, minor additive noise and
compression, although it suffers from changes in brightness and large geometric
transforms. Furthermore, Han et al [21] and Wu et al [41] presented an in-depth
21
review and analysis on content based image authentication; whereas Kim et al [42]
proposed a novel scheme to detect unauthorized copies of an image using DCT
coefficients.
Venkatesan et al [43] explained a novel image indexing technique called โimage hash
functionโ. This algorithm uses randomized strategies for a non-reversible
compression of images into random binary strings and is robust against limited
attacks. Lefรจbvre et al[44] employ random transform for feature extraction and
principle component analysis to reduce the hash length, nevertheless, its robustness
for texture image is limited. Additionally, Swaminathan et al [45] advocated an
algorithm for image hashing based on Fourier transformed controlled randomization,
though, the algorithm suffers from known attacks, for instance additive noise.
Khelifi et al [46] introduced virtual watermark detection using an optimum
multiplicative watermark detector. It uses a pseudo-randomly generated pattern to
extract the hash bits. These researchers [47] also presented an analysis of the security
of a perceptual image hash based on non-negative matrix factorization and reveal
that the use of a secret key combined with image dependent keys can enhance
security. Moreover, Kozat et al [48] explain singular value decomposition for image
hashing; however, this algorithm, although it is robust is limited.
Recently, Monga et al [14] suggested an image hashing algorithm using visually
significant feature points and performed a performance evaluation and tradeoffs
between geometric invariance and robustness against classical attacks.
Also, Lv et al [16] created new image hashing algorithm using a local feature point
with SIFT to detect robust feature points and incorporate Harris criterion to select the
most stable points that are less vulnerable to image processing attacks.
More recently, Badrinath et al [49] advocated an efficient indexing scheme for a
palmprint identification system, which makes use of a fusion of votes obtained
through strategy-based geometric hashing and a SURF (speeded-up robust feature)
score. Moreover, Tuytelaars et al [50] presented an overview of feature detector,
such as corner (via Harris, SUSAN: Smallest Univalue Segment Assimilating
22
Nucleus), Blob (via Hessian) and a salient region (via MSER: Maximally Stable
Extremal Regions). The overview exposes the property of feature points and their
invariant under significant geometric transforms.
Lowe D.G [51] illustrated a method for extracting distinctive invariant features from
images which can be used to perform reliable matching between different views of
an object. The algorithm used is SIFT, as it transforms image data into scale
invariant coordinates relative to local features.
Ahmed et al [13] recommended a hash-based image authentication scheme, which
concentrates on various issues like tamper detection, security and robustness. A
secret key is used at the feature extraction stage to randomly modulate image pixels,
in order to create a transformed feature space. The hash based scheme offers good
robustness against JPEG compression, and low and high-pass filtering. In addition,
Koval et al [52] proposed two classes of robust-hashing techniques: Random-Based
Hashing and Content-Based Hashing systems, and perform security analysis for each
class to demonstrate how security issues arise.
2.3 Template Protection Techniques
Template protection is a collective term for a variety of methods that aim to preserve
privacy and enhance the secure storage of biometric data. Jain et al [7] presented an
overview of biometric template protection schemes and categorize them into a
transformation-based approach and biometric cryptosystems [7] [8] as shown in
Figure 2.1. The functions used in transformation approaches can distort or randomize
biometric data so that the original data cannot be reconstructed from transformed
templates. The biometric cryptosystems can be embedded or generate secrets from
the biometric data.
The biometric template protection scheme has the following properties:
Diversity: where the secure template must not allow cross-matching
across the database, by ensuring the privacy of the user.
23
Revocability: revokes a compromised template and reissues a new one
based on the same biometric data.
Security: prevents an adversary from creating a physical spoof of the
biometric trait from a stolen template.
Performance: the biometric template scheme should not degrade the
recognition performance of the (True Positive Rate and False Positive
Rate) biometric system.
Template Protection
Feature Transformation Biometric Cryptosystem
Salting
(e.g., Biohashing)
Non- invertible
Transform
(e.g., Robust Hashing
Key Generation
(e.g., Secure Sketch Fuzzy
Extractor)
Key Binding
(e.g., Fuzzy Vault, Fuzzy
Commitment)
Figure 2.1: Template protection Schemes [5].
Ratha et al [9] outlined the inherent strengths of biometrics-based authentication in
identifying weak links in systems by pointing to possible attacks in a generic
biometric system. Furthermore, Nagar et al[11] explained that the matching
performance and security of a fingerprint fuzzy vault is improved by incorporating a
minutiae descriptor. However, Tuyls et al [53] proposed template protection for
fingerprint based authentication, where the algorithm is based on helper data
consisting of two parts. The first part identifies the reliable components with a high
24
signal to noise ratio in the Gabor-filtered fingerprint, whilst the second part allows
for noise correction on the quantized representation.
Breebaart et al [54] and Rane [55] described the requirements for the standardization
of architecture in biometric data storage and processing, which satisfy the conditions
of renewability (are properties that allow a system to recover after the auxiliary data
and/or the pseudonymous data have been compromised by an attacker),
irreversibility, unlinkability (are the property possessed by two or more biometric
references between auxiliary data and pseudonymous data pairs by virtue of which
they cannot be linked to each other or to the individual from whom they were
derived), confidentiality and integrity. Sun et al [56] presented a key-mixed template,
which mixes a userโs template with a secret key to generate another form of template,
in order to prevent the biometric template stored in the database from experiencing
backend attacks, snooping and tampering assaults.
In addition, Roberge et al[57] proposed a biometric encryption algorithm for the
connection and retrieval of digital keys, which can be used as a method in the secure
management of cryptographic keys. Sutcu et al [58] examine the storage of a face
biometric template by applying a secure sketch algorithm, and noted the performance
and the security of the secure sketch method. In addition, Brindha V.E [59] attempts
to improve template security by combining a dorsal hand vein biometric with
cryptography to generate a fuzzy vault. Sutcu et al [60] proposed a geometric
transformation for securing minutiae based fingerprint templates. This method is a
robust one-way transformation that maps the geometrical configuration of the
minutiae points into a fixed-length code vector.
Mirmohamadsadeghi et al [61] established a new technique to protect fingerprint
minutiae based on a MCC. This is a hybrid technique, combining a transformation
and a user key, which provides diversity, revocability, and irreversibility for MCC
descriptors with respect to the original minutiae of the fingerprint image. In addition,
Prasad et al [62] described an alignment free method for generating the cancellable
template, which use neighbouring relations around every reference minutiae.
25
Finally, Jain et al[63] provide a theoretical framework, which includes template
security requirements, protection approaches and various fingerprint template
protection schemes in detail. Table 2.1 explains the different techniques to transform
fingerprint features for template protection schemes.
Table 2.1: Different methods used to transform fingerprint features for template
protection [63].
Technique Features Transformation Final
Representation
Spectral minutiae [29] Minutiae Fourier transform of 2D-
delta functionsat minutiae
locations
Vector
BioPhasor [64] FingerCode Nonlinear Vector
Biometric encryption [65] Fingerprint image Apply a secure filter Vector
Minutiae indicator [66] Minutiae Minutiae locations marked
as โ1โ
Vector
Histogram of minutiae
triplets [67]
Minutiae Hashing the histogram of
minutiae triplet features
Vector
Cuboid based minutiae
Aggregates [68]
Minutiae Minutiae aggregate
selection from random
local regions
Vector
Symmetric hash [26] Minutiae as complex
numbers
Set of order invariant
functions for minutiae
Minutiae
Cancelable fingerprints [67] Minutiae Image folding Minutiae
Alignment free cancellable
fingerprint [34]
Minutiae orientation
field
Transform minutiae
according to surrounding
orientation field
Minutiae
Minutiae structures [68] Minutiae Local minutiae structures Minutiae
26
2.4 Summary
In this chapter, a brief review on securing template protection and its technique are
discussed in detail. In relation to the proposed research work, the literature was
carried out on three main categories, which were discussed: minutiae based
fingerprint templates, feature extraction image hashing and template protection. This
research focuses on improving the security of fingerprint templates to allow accurate
comparison of the fingerprint content. The current methods to generate fingerprint
templates for comparison purposes mostly rely on using a single feature extraction
technique such as SIFT or Fingerprint Minutiae. However, the combination of two
feature extraction techniques (e.g., SIFT-Minutiae) has not been studied in the
literature. The results in this thesis demonstrate that new approach contributes
towards the improvement of the fingerprint template and accuracy of fingerprint
matching improved when combinations of two different feature extraction techniques
are used, in contrast to using only a single feature extraction technique.
In the next chapter, the basic characteristics of transform domain techniques are
presented. These techniques are based on the manipulation of the orthogonal
transform of image rather than the image itself. Transform domain techniques are
suited for processing the image according to the frequency content. The principle
behind this domain methods of image enhancement consists of the computing a 2-D
discrete unitary transform of the image, for instance the 2-D DFT manipulating the
transform coefficients by an mathematical operator and then performing the inverse
transform. The orthogonal transform of the image has two components magnitude
and phase. The magnitude consists of the frequency content of the image. The phase
is used to restore the image back to the spatial domain. The usual transform domain
enables operation on the frequency content of the image, and therefore high
frequency content such as edges and other subtle information can easily be enhanced.
27
Chapter 3
Image Representation in the Transform Domain
With the increasing demand of enhanced security in our daily lives, reliable personal
identification through biometric is currently active research. An individual can be
identified effectively using biometric modalities such as fingerprint, palm-print and
face. Recently many researchers have proposed several promising methods for
biometric image identification using transform domains: Discrete Fourier Transform
(DFT), Discrete Cosine Transform (DCT) and Fourier-Mellin Transform (FMT) and
Discrete Wavelet Transform (DWT).
Kapil Rathor [71] proposed a method to improve the recognition rates of iris images
by localizing the image between the inner and the collaretee boundary. 2-D DFT is
used for finding the collaretee boundary. Imtiaz et al [72] presented a spectral feature
extraction algorithm for palmprint recognition. In this method, the entire image is
segmented into several spatial modules and the task of the feature extraction is
carried out using 2-D Fourier transform within those spatial modules and shown
good recognition accuracy and computational complexity.
Amornraksa et al [73] presented a fingerprint recognition method based on the DCT
features. Applying the DCT transform to a discrete fingerprint image, the DCT
features used for fingerprint matching resulting in higher recognition rates and a
lower complexity. Imtiaz et al [74] proposed a DCT-based palm-print recognition
scheme, where the dominant spectral features are extracted separately from each of
the narrow-width band resulting from image segmentation operation. These feature
extraction scheme offers two advantages: first, it captures local variations that exist
in the palmprint images, which plays an important role in discriminating different
persons. Second, it utilizes a very low dimensional feature space for the recognition
task, which ensures lower computational burden. Badrinath et al [75] extracted the
feature using 1-D DCT coefficients to design an efficient palm print based
recognition system.
Singh et al [76] proposed a novel rotation-invariant and degraded partial palmprint
recognition method, which combines the features of the FMT and Modified Phase โ
28
Only correlation. Additionally, Prungsinchai et al [77] presented an efficient secure
and robust perceptual image hashing technique based on the Fourier-Mellin
Transform. It has been shown this method is robust against signal processing
operation and geometric attacks. It has also been shown that FMT based features
outperform SVD, wavelet and NMF based hashing under geometric distortions.
Ekinci et al [78] presented a novel Daubechies-based kernel Principal Component
Analysis (PCA) method by integrating the Daubechies wavelet representation of
palm images and the kernel PCA method for palmprint recognition. Yang et al [79]
combine fingerprint, palmprint and hand geometry for person identity verification.
In this multimodal system, wavelet transform is employed to extract feature from
fingerprint, palmprint and hand-geometry. The Feature fusion and match score
together are used for identification.
In the following section, the most widely used transform domain techniques are
reviewed.
3.1 Discrete Fourier Transform (DFT)
The Fourier Transform is an important image processing tool which is used to
decompose an image into its sine and cosine components. The Fourier transform
of ๐ฅ ๐ก is defined as
๐ ๐ = ๐ฅ ๐ก +โ
โโ
๐โ๐2๐๐๐ก ๐๐ก
(3.1)
Where the independent variable ๐ก represents time, the transform variable ๐
represents ordinary frequency. In the Fourier domain signal, each point represents
a particular frequency contained in the time domain signal. The signal ๐ฅ(๐ก) can
be reconstructed from X(f) by the inverse transform
๐ฅ ๐ก = ๐ ๐ +โ
โโ
๐โ๐2๐๐๐ก ๐๐
(3.2)
29
The interpretation of ๐ ๐ is aided by expressing it in polar coordinate form as
๐ ๐ = ๐ ๐ ๐๐ฮฆ(๐) (3.3)
Where ๐ ๐ and ฮฆ(๐) represent the amplitude and the phase of ๐ ๐ ,
respectively.
Let ๐ฅ(๐ก) โ ๐ ๐ are a Fourier transform pair. Some important properties of the
Fourier transform are
Linearity
๐๐ฅ1 ๐ก + ๐๐ฅ1 ๐ก โ ๐๐1 ๐ + ๐๐2 ๐ (3.4)
Convolution
๐ฅ1 ๐ก โ ๐ฅ2 ๐ก โ ๐1 ๐ ๐2 ๐ (3.5)
Scaling
๐ฅ ๐๐ก โ 1
๐ ๐
๐
๐ (3.6)
Modulation
๐ฅ ๐ก ๐โ๐2๐๐0๐ก โ ๐ ๐ โ ๐0 (3.7)
Parsevalโs theorem
โซโ ๐ฅ(๐ก) 2 = โซ
โ ๐ ๐ 2 (3.8)
The DFT is the sampled Fourier Transform and therefore, it does not contain all
frequencies forming an image, but only a set of samples that are large enough to
perfectly describe the spatial domain image. The number of frequencies
corresponds to the number of pixels in the spatial domain image, i.e. the image in
the spatial and Fourier domain are of the same size [80].
For a square image size N x N, the two-dimensional DFT is given by
๐น ๐, ๐ = ๐ ๐, ๐ ๐โ๐2๐
๐๐
๐+
๐๐
๐
๐โ1
๐ =0
๐โ1
๐=0
(3.9)
where = 0 . . . . . N โ 1 ; ๐ = 0 . . . . . M โ 1
30
๐(๐, ๐) is the image in the spatial domain and the exponential term is the basis func-
tion corresponding to each point ๐น ๐, ๐ in the Fourier space. Equation (3.9) can be
interpreted as: the value of each point ๐น ๐, ๐ is obtained by multiplying the spatial
image with the corresponding base function and summing up the result. The basis
functions are sine and cosine waves with increasing frequencies, i.e. F(0,0)
represents the DC-component of the image which corresponds to the average
brightness and F(N-1, M-1) represents the highest frequency.
Similarly, the inverse Fourier transform is given by
๐ ๐, ๐ =1
๐๐ ๐ ๐, ๐ ๐
๐2๐ ๐๐
๐+
๐๐
๐
๐โ1
๐=0
๐โ1
๐=0
(3.10)
where ๐ = 0 . . . . . N โ 1, ๐ = 0 . . . . . M โ 1
3.2 Fourier-Mellin Transform (FMT)
The Fourier-Mellin transform is a important tool for pattern recognition,
reconstruction and image database retrieval, because its resulting spectrum is
invariant in rotation, translation and scaling .
Let ๐ denote a function representing a grey level image defined over a set of โ2. The
standard Fourier-Mellin transform of ๐ is given by
โ ๐, ๐ฃ โ โ ร โ, ๐๐ ๐, ๐ฃ =1
2๐ ๐ ๐, ๐
2๐
0
โ
0
๐โ๐๐ฃ๐โ๐๐๐ d๐๐๐
๐
(3.11)
where ๐ ๐, ๐ is the polar coordinates representation of a 2-D function.
๐ is assumed to be summable over โ+โ ร ๐1 under the measure d๐
๐๐
๐, i. e.
๐ ๐, ๐ ๐โ๐๐ฃ๐โ๐๐๐ d๐๐๐
๐
2๐
0
โ
0
= 1
๐๐ ๐, ๐ d๐๐๐ < โ
2๐
0
โ
0
(3.12)
31
Since ๐ is positive. For โ ๐, ๐ฃ โ โ ร โ. โ1 denotes the additive group of integers,
โ denotes the additive group of the real line. โ+โ the multiplicative group of positive
and nonzero numbers, ๐1 the unit circle of the plane โ2 . All these groups are
locally compact and commutative. The direct product โ+โ ร S1 forms a locally
compact and commutative group under the following law: ๐ผ, ๐ ฮฟ ฯ, ฯ = (๐ผฯ, ๐ +
ฯ)[81].
The FMT could be divided into main three steps, which result in the invariance to
rotation, scaling and translation attacks:
The Fourier Transform (FT): It converts the original image in spatial domain onto
spectrum domain. The magnitude of Fourier transform itself is the translation in-
variant.
The Cartesian to Log-Polar Coordinates: The conversion to log-polar coordinates
converts the scale and rotation differences to vertical and horizontal offsets that
can be measured.
The Mellin Transform: A second FT, called the Mellin Transform (MT) gives a
transform-space image that is invariant to rotation, scaling and translation.
3.3 Discrete Cosine Transform (DCT)
The discrete cosine transform attempts to decorrelate the image data. After
decorrelation each transform coefficient can be encoded independently without
losing compression efficiency [82].
The most common DCT definition of a 1-D sequence of length N is
๐ถ ๐ข = ๐ผ ๐ข ๐ ๐ฅ cos
๐โ1
๐ฅ=0
๐ 2๐ฅ + 1 ๐ข
2๐
(3.13)
For u= 0, 1,โฆ..N-1. Similarly, the inverse transformation is defined as
32
๐ x = ๐ผ ๐ข ๐ถ ๐ข cos
๐โ1
๐ข=0
๐ 2๐ฅ + 1 ๐ข
2๐
(3.14)
For ๐ฅ = 0, 1,โฆ..N-1. In both equation (3.13) and (3.14) ๐ผ ๐ข is defined as
๐ผ ๐ข =
1
๐ ๐๐๐ ๐ข = 0
2
๐ ๐๐๐ ๐ข โ 0
(3.15)
The 2-D DCT is a direct extension of the 1-D case and is given by
๐ถ ๐ข, ๐ฃ = ๐ผ(๐ข)๐ผ(๐ฃ) ๐ x, y
๐โ1
๐ฆ=0
๐โ1
๐ฅ=0
cos(๐ 2๐ + 1 ๐ข
2๐)cos(
๐ 2๐ + 1 ๐ฃ
2๐)
(3.16)
For ๐ข, ๐ฃ = 0, 1, 2โฆ..N-1. The inverse transform is defined by
๐ x, y = ๐ผ ๐ข ๐ผ ๐ฃ ๐ถ ๐ข, ๐ฃ
๐โ1
๐ฃ=0
๐โ1
๐ข=0
cos ๐ 2๐ + 1 ๐ข
2๐ cos
๐ 2๐ + 1 ๐๐ฃ
2๐
(3.17)
For x, y = 0, 1, 2โฆ..N-1. The main advantages of the DCT are that it gives a real
output image and that it is a fast transform. A major use of the DCT is in image
compression. Indeed, after performing a DCT it is possible to discard the coefficients
representing high frequency components that the human eye is not very sensitive to. Thus,
the amount of data can be reduced; without seriously affecting the way an image appears to
the human eye.
3.4 Discrete Wavelet Transform (DWT)
The wavelet decomposition is a mathematical tool allowing the study of signals and
signal generating processes characterised by a non-stationary behaviour. It accounts
for the evolution in time of the frequency content of a signal [80]. A signal ๐ฅ ๐ก can
33
often be better analysed, described, or processed if expressed as a linear
decomposition by
๐ฅ ๐ก = ๐๐ ,๐
๐ ,๐
2๐
2๐(2๐ ๐ก โ ๐)
(3.18)
Where the two-dimensional set of coefficients ๐๐ ,๐ is called discrete wavelet
transform (DWT) of ๐ฅ ๐ก . Note that the basis functions ๐๐ ,๐ (๐ก) = 2๐
2๐(2๐ ๐ก โ ๐) are
generated from a single function ๐ ๐ก called โmother waveletโ by changing two
parameters j and k. The location of the wavelet moves in time or space, as the
index k changes. This allows the expansion to explicitly represent the
location of events in time or space and enables a representation of detail or
resolution. A more precise way of indicating how the ๐๐ ,๐ โฒs are calculated can
be written using the inner products as
๐ฅ ๐ก = ๐๐ ,๐ ๐ก , ๐ฅ ๐ก ๐๐ ,๐ ๐ก ๐ ,๐ (3.19)
3.4.1 Multiresolution Analysis
The multiresolution formulation of wavelet systems is designed to represent signals
where a single event is decomposed into finer and finer details. As described earlier
for the wavelet, a set of scaling function is defined in terms of integer translated of
the basic scaling function ๐(t) by
๐๐=๐ ๐ก โ ๐ ; ๐ โ ๐, ๐ โ ๐ฟ2 (3.20)
`The subspace of ๐ฟ๐โ spanned by these function is defined as
๐0 = ๐๐๐๐๐ (๐๐(๐ก)) (3.21)
A two-dimensional family of functions is generated from the basic scaling function
by scaling and translation by
๐๐ ,๐ (t)=2๐ /2๐(2๐ ๐ก โ ๐) (3.22)
whose span over ๐ is
34
๐๐ =๐๐๐๐๐(๐๐(๐ก)) (3.23)
This means that if ๐ฅ ๐ก โ ๐๐ , then it can be expressed as
๐ฅ ๐ก = ๐๐
๐
๐(2๐ ๐ก โ ๐)
(3.24)
For j>0, the span can be larger since ๐๐ ,๐ is narrower and is translated in smaller
steps, the basic requirement of multi resolution analysis is
๐0 โ ๐1 โ ๐2 โฆ โฆ . . โ ๐ฟ2 (3.25)
Hence, the spaces ๐๐ satisfy a natural scaling condition
๐ฅ(๐ก) โ ๐๐ โ ๐ฅ(2๐ก) โ ๐๐ +1 (3.26)
The important features of a signal can be better described by also using a set of
wavelet functions ๐๐ ,๐ (t) that span the differences between the successive spaces๐๐ .
Let us denote the orthogonal complement of ๐๐ in ๐๐ +1 as ๐๐ . It follows
๐1=๐0 + ๐0 (3.27)
Which extends to
๐๐= ๐0 + ๐0 + ๐1 + โฏ + ๐๐โ1 (3.28)
Therefore, a signal ๐ฅ(๐ก) โ ๐๐ can be expressed as
๐ฅ ๐ก = ๐๐
๐
๐ ๐ก โ ๐ + ๐(๐, ๐)
๐
๐โ1
๐=0
๐๐ ,๐(t)
(3.29)
Equation (3.29) represents a decomposition of ๐ฅ ๐ก with n resolutions (or scales).
The first summation gives a function that is a low resolution or coarse approximation
of ๐ฅ ๐ก . For each increasing index j in the second summation, a higher or finer
resolution is added, which adds increasing detail. This is somewhat analogous to a
Fourier series where the higher frequency terms contain the detail of the signal. From
Equation (3.25) it can be observed that if a function ๐ ๐ก is in ๐ฃ๐โ1 , it is also in ๐๐ ,
35
which is the space spanned by ๐(2๐ ๐ก). This means ๐(2๐โ1๐ก) can be expressed in
terms of a weighted sum of shifted ๐(2๐ ๐ก)
๐(2๐โ1๐ก) = ๐(๐)
๐
2๐ /2๐(2๐ ๐ก โ ๐)
(3.30)
Similarly, since ๐๐ โ1 โ ๐๐ , ๐(2๐โ1๐ก) can be expressed as
๐(2๐โ1๐ก) = ๐(๐)
๐
2๐/2๐(2๐ ๐ก โ ๐)
(3.31)
Assume a signal ๐ฅ(๐ก) โ ๐๐ which can therefore be written as
๐ฅ ๐ก = ๐๐โ1,๐2๐โ1
2 ๐ 2๐โ1๐ก โ ๐ +
๐
๐๐โ1,๐ 2๐โ1
2 ๐ 2๐โ1๐ก โ ๐
๐
(3.32)
where
๐๐โ1,๐ = ๐ฅ ๐ก , 2๐โ1
2 ๐ 2๐โ1๐ก โ ๐ = ๐ฅ ๐ก , 2๐โ1
2 ๐ 2๐โ1๐ก โ ๐ โ
โโ
๐๐ก
(3.33)
and
๐๐โ1,๐ = ๐ฅ ๐ก , 2๐โ1
2 ๐ 2๐โ1๐ก โ ๐ = ๐ฅ ๐ก , 2๐โ1
2 ๐ 2๐โ1๐ก โ ๐ โ
โโ
๐๐ก
(3.34)
From equation (3.30) and (3.32) one can deduce
๐๐โ1,๐ = ๐(๐ โ 2๐)
๐
๐๐ ,๐
and (3.35)
๐๐โ1,๐ = ๐(๐ โ 2๐)๐ ๐๐ ,๐
(3.36)
36
The last two equations represent a digital filtering process followed by a down
sampling (also called decimating) by a factor of 2. The down sampling takes a signal
๐ฅ ๐ as an input and produces an output ๐ = ๐ฅ 2๐ . These equations shows that
the scaling and wavelet coefficients at different levels of scale can be obtained by
convolving the expansion coefficients at scale j by the time reversed recursive
coefficients ๐ โ๐ and ๐ โ๐ then down sampling to give the expansion
coefficients at the next level of j-1. In other words, the scale j coefficients are filtered
by two FIR digital filters with coefficients ๐ โ๐ and ๐ โ๐ . Subsequently, the
down โsampling gives the next coarser scaling and wavelet coefficients. These
structures implement Mallatโs algorithm [83] and have been developed in filter bank,
quadrature mirror filters, conjugate filters, and perfect reconstruction filter bank in
the literature [84]. Mallat, Daubechies, and others showed the relation of wavelet
coefficient calculation and filter banks. The implementation of Equation (3.35) and
(3.36) is illustrated in Figure 3.1. where the down โpointing arrows denote a down
sampling by two and other boxes denote convolution by ๐ โ๐ or๐ โ๐ . This
splitting, filtering, and decimation can be repeated on the scaling coefficients to give
the two-scale structure in Figure 3.2.
2h(-n)
g(-n)
aj
aj-1
2 dj-1
Figure 3.1: One-stage wavelet decomposition
2h(-n)
g(-n)
2
h(-n)
g(-n)
2
2 aj
aj-1
dj-1
aj-2
dj-2
Figure 3.2: Two-stage wavelet decomposition
37
This Splitting, filtering, and decimation can be repeated on the original fine scale
coefficients and can be made from a combination of the scale function and wavelet
coefficients at a coarse resolution. This derived by considering a signal in the j
scaling function space ๐ฅ(๐ก) โ ๐๐ , which can be expressed as given by Equation (3.32)
or in terms of the scaling function at the same level j by
๐ฅ ๐ก = ๐๐ ,๐
๐
2๐ /2 ๐(2๐ ๐ก โ ๐)
(3.37)
Substituting Equation 3.30 and 3.31into 3.32 gives
๐ฅ ๐ก = ๐๐โ1,๐
๐
๐(๐)
๐
2๐
2๐ 2๐ ๐ก โ 2๐ โ ๐
+ ๐๐โ1,๐
๐
๐(๐)
๐
2๐
2๐ 2๐ ๐ก โ 2๐ โ ๐
(3.38)
Because all of these functions are orthogonal, multiplying the equation 3.37 and 3.38
by ๐ 2๐ ๐ก โ ๐โฒ
By integrating the coefficient can be rewritten as:
๐๐ ,๐ = ๐๐โ1,๐
๐
๐ ๐ โ 2๐ + ๐๐โ1,๐
๐
๐ ๐ โ 2๐
(3.39)
The final equation is actually evaluated by up-sampling the (j-1) scale coefficient
sequence ๐๐โ1,๐ , which means double its length by inserting zeros between each
term, then convolving it with the scaling filter ๐ ๐ . The same procedure is
performed to the (j-1) level wavelet coefficient sequence ๐๐โ1,๐ and the results are
added to produce the j level scaling function coefficients ๐๐ ,๐ .This structure is
illustrated in Figure3.3. This process can be continued to any level by combining the
appropriate scale wavelet coefficients.
38
2
h(n)
g(n)
+
2
aj-1
dj-1
aj
Figure 3.3: One-stage wavelet reconstruction
The resulting two-scale tree is shown in Figure 3.4
2
h(n)
g(n)
+ h(n)
g(n)
+
2
2
2
aj-2
aj-1
aj
dj-2
dj-1
Figure 3.4: Two-stage wavelet reconstruction.
3.4.2 Properties
In practical application, wave bases can judiciously be chosen to fit the
behaviour of the data to be analysed. An excellent choice of the wavelet bases
can optimise signal processing algorithms. Indeed, a wavelet basis that produces
more coefficients with a magnitude closed to zero is preferred more in data
compression, since theses coefficients require less bits to encode. The most
relevant criteria are the number of vanishing moments, the size of the support
and regularity.
The number of vanishing moments is related to the smoothness or
differentiability of ๐(๐ก) and (๐ก). The size of the support measure the interval in
39
time in which the wavelet takes non-zero values. Regularity is defined ill terms
of zeros of the frequency response function of the scaling filter ๐(๐) thus,
indicating how fast the Fourier transform magnitude drops off, as the frequency
progresses to infinity. This is particularly related to the frequency localisation of
the decomposed signal.
The size of the wavelet support increases with the number of vanishing
moments. The wavelet regularity is important to reduce the artefacts. The choice
of an optimal wavelet in image compression is thus the result, of a trade-off
between the number of vanishing moments and artefacts. Some useful properties
of the wavelet transform can be summarised as follows:
They can represent smooth functions
They can represent singularities
The basis functions are local. This makes most coefficient-based algorithms
naturally adaptive to inhomogeneities in the function.
They have the unconditional basis property for a variety of function classes
implying that if one does not know much about a signal (for instance, a
signal with a non-stationary behaviour), the wavelet basis is usually a
reasonable choice.
They are computationally inexpensive with a complexity O (N) compared to a
Fourier transform, which is Nlog(N) or an arbitrary linear transform
which is O(N2).
3.4.3 2-D wavelet transform
For 2-D data such as images, the most commonly used algorithm for wavelet
decomposition uses separable one-dimensional wavelets and scaling functions.
This kind of two-dimensional DWT leads to a decomposition of approximation
coefficients at level j in four components: the approximation at level
40
jโ1(๐๐โ1), and the details in three orientations (horizontal๐๐โ1(๐)
, vertical๐๐โ1(๐ฃ)
,
and diagonal๐๐โ1(๐)
), Figure 3.5 describes the basic decomposed steps for images.
An example of a one stage decomposed image of โLenaโ is illustrated by
Figure3.6. In a similar way, the reverse process can be used to obtain the
original 2-D signal.
Figure 3.5: One-stage 2-D wavelet decomposition
(a) Original Image (b) Decomposed Image
Figure 3.6: One-stage 2-D wavelet decomposition of โLenaโ.
3.5 Summary
In this chapter, the basic characteristics of transform domain: Discrete Fourier
Transform, Discrete Cosine Transform and Fourier-Mellin Transform and Discrete
Wavelet Transform, are reviewed. These techniques are very promising, within
biometric and image processing. The transform domain technique are the most
challenging part of the image hashing in the feature extraction stage, the extracted
41
feature are invariant to image processing operation and geometric attacks. In next
chapter, the recent state-of-the-art feature extraction techniques such as: end-stopped
wavelets, SURF, SIFT, and SIFT-Harris are presented.
42
Chapter 4
Image Feature Extraction Techniques
4.1 Introduction
Recent research in biometric recognition systems has been performed to secure the
biometric through hash based techniques. An image hash can be constructed by a set
of features from images to form a compact representation that can be used for the
authentication and integrity of the data. The selection of a feature extraction
technique is a key step in any biometric recognition system and all pattern
recognition systems. The recognition process analyses the spatial geometry of the
distinguishing feature of the image. Different methods exist to extract the identifying
features of an image, although in general they can be classified into three
approaches: Feature-based approaches, Appearance-based approaches and Hybrid
approaches
Feature-based approaches: This approach is based on the properties of
individual appendages located on a biometric trait, such as eyes, nose and
mouth on a face, wrinkles lines for a palm print, eye lashes for an iris, etc, as
well as their relationships with each other.
Appearance-based approaches: These are based on information theory
concepts and seek a computational model that describes a biometric image.
This works by extracting the most relevant information contained in the
image without dealing with the individual appendages.
Hybrid approaches: This approach uses both holistic feature and local
features. Modular eigenfaces and component-based feature can be given as
examples.
Once a raw image is properly pre-processed (i.e., enhancement, formatting,
segmentation, etc), the feature extraction algorithm can then be used to extract the
relevant features. The feature extraction algorithms can be classified into two groups:
Global Feature Extractors and Local Feature Extractors.
43
Global Feature Extractors: Aim to locate usable features from the raw data
at the overall image level. They process the image as a whole and attempt to
extract the features, e.g. the Gabor wavelet-based approach for iris and
fingerprint recognition.
Local Feature Extractors: Focus on the block of image data. These
algorithms work on small windows within the images and extract the relevant
features, e.g. minutiae extraction.
The feature points should be largely invariant under perceptually insignificant
distortions. The feature extraction techniques discussed in terms of robustness to
content-preserving operations include rotation, translation and other variations. The
state of the art feature extraction techniques: End Stopped Wavelets [14], SIFT [51],
SURF [15] and SIFT-Harris [16] are approached to solve problems in the image
based biometric, including fingerprint, face, etc.
4.2 End-Stopped Wavelets
Psychophysical studies have identified the presence of certain cells, called
hypercomplex or end stopped cells, in the primary visual cortex. Two types of end-
stopped cells have being identified. The single end-stopped cells respond strongly to
extremely robust image features, for instance corner like stimuli and points of high
curvature. The second type of end-stopped cells responds strongly to a linear
segments or curved lines. The term end-stopped comes from the impressive
sensitivity of these cells to end-points of linear structures, and moreover,
Bhattacharjee et al [85] construct โโend โstoppedโโ wavelets to capture this
behaviour. The construction of the wavelet kernel or basis function combines two
operations. Firstly, linear structures having a certain orientation are selected. These
linear structures are then processed to detect line-ends (corners) and or high
curvature points. Morlet wavelets can be used to detect linear structures having a
specific orientation. In the spatial domain, the 2-D Morlet wavelet is given as [85];
๐๐(๐) = ๐ jk0.X โ ๐โ1
2 k0 2
๐โ1
2 X 2
(4.1)
44
where X = (๐ฅ, ๐ฆ) represents 2-D spatial coordinates, and ๐พ0 = (๐พ0, ๐พ1) is the wave-
vector of the mother wavelet, which determines the scale-resolving power and
angular-resolving power of the wavelet. The frequency domain
representation ๐๐(๐พ) of a Morlet wavelet is
๐ ๐(๐พ) = ๐โ1
2 kโk0 2
โ ๐โ1
2 k0 2
๐โ1
2 k 2
(4.2)
Here, ๐พ represents the 2-D frequency variable (๐ฐ, ๐ฑ). The Morlet function is similar
to the Gabor function, although with an extra correction term ๐โ1
2 k 2+ X 2
to make
it an admissible wavelet. The orientation of the wave-vector determines the
orientation tuning of the filter. A Morlet wavelet detects linear structures orientated
perpendicular to the orientation of the wavelet.
In two dimensions, the end points of the linear structures can be detected by applying
the first derivative of the Gaussian (FDoG) filter in the direction parallel to the
orientation of the structures in question. The first filtering stage detects lines having a
specific orientation and the second filtering stage detects end-points of those
particular lines. These two stages can be combined into a single filter to form an
โend-stoppedโ wavelet (Figure 4.1). For example, the end-stopped wavelet and its 2-
D Fourier transform is as follows:
๐๐ธ(๐ฅ, ๐ฆ) = 1
4๐ฆ๐โ(
๐ฅ2+๐ฆ 2
1+
๐0
1 ๐0โ2๐๐ฅ ) (4.3)
๐ ๐ธ(๐ฐ, ๐ฑ) = 2ฯ ๐โ ๐ฐ2โ๐0 +(๐ฃ2)
2 ๐๐ฃ๐โ ๐ฐ2+๐ฃ2
2 (4.4)
Equation (4.4) shows ๐ ๐ธ as a product of two factors. The first factor is a Morlet
wavelet orientated along the axis. The second factor is a FDoG operator applied
along the frequency axis ๐ฑ, i.e., in a perpendicular direction to the Morlet wavelet.
Hence, this wavelet detects line ends and high curvature points in the vertical
direction.
45
(a) (b) (c)
Figure 4.1: Behaviour of the end-stopped wavelet on a synthetic image. (a) Synthetic
L-shaped image. (b) Response of a Morlet wavelet. (c) Response of the end-stopped
wavelet[14].
Monga et al [14] recommend a feature extraction technique that computes a wavelet
transform based on an end-stopped wavelet obtained by applying the FDoG operator
to the Morlet wavelet
๐๐ธ ๐ฅ, ๐ฆ, ๐ = ๐น๐ท๐๐บ ๐ ๐๐ ๐ฅ, ๐ฆ, ๐ (4.5)
and orientation tuning is given by ๐ = tanโ1((๐1 ๐0 )).
The orientation range [0,๐] need be divided into M intervals and the scale parameter
๐ผ be sampled exponentially as ๐ผ๐ , ๐ โ ๐. This results in a wavelet family
(๐๐ธ ๐ผ๐ ๐ฅ, ๐ฆ, ๐๐ , ๐ผ โ โ, ๐ โ ๐ (4.6)
Where ๐๐ = ๐๐ ๐ , ๐ = 0, โฆ โฆ๐ โ 1.
The wavelet transform is
๐๐ ๐ฅ, ๐ฆ, ๐ = ๐(๐ฅ1, ๐ฆ1) ๐๐ธโ ร ๐ผ๐ ๐ฅ โ ๐ฅ1, ๐ฆ โ ๐ฆ1 , ๐ ๐๐ฅ1๐๐ฆ1 (4.7)
Mongaโs feature detection method preserves significant image geometry feature
points of an image as: (1) Computes the wavelet transform for each image location
(2) Identifies significant features by looking for local maxima of the magnitude of
the wavelet coefficients in a preselected neighbourhood. (3) Thresholding to
eliminate spurious local maxima in featureless region of the image. These feature
detection methods have two free parameters: integer scale ๐ and real threshold T.
The threshold is used to select a fixed number of feature points from the image and
46
an image feature vector is formed by collecting the magnitudes of the wavelet
coefficients at the selected feature points.
4.2.1 Experimental Results
In this novel work, Hausdorff distance is used to characterize the robustness and
discriminability of the feature points.
a) Hausdorff Distance [87]
Given two finite points sets A= ๐1, ๐2, ๐3, ๐4 โฆโฆ . . ๐๐ and B= ๐1, ๐2, ๐3, ๐4 โฆโฆ . . ๐๐ ,
the Hausdorff distance is defined as
H (A, B) = max (h (A, B), h (B, A)) (4.8)
Where h (A, B) = max๐โ๐ด min๐โ๐ต ๐ด โ ๐ต (4.9)
and . is the underlying norm on the points of A and B. The function h (A, B) is
called the directed Hausdorff distance from A to B. h (A,B) in effect rank each point
of A based on its distance to the nearest point on B and subsequently uses the largest
ranked point as the distance. The Hausdorff distance H (A, B) is the maximum of
h (A, B) and h (B, A).
b) Result and Analysis
The algorithm is tested on a fingerprint image (Figure 4.2) database of 100 images
from FVC2004/DB1_A [88]. The similarity of the feature point hash to the original
and identical images are evaluated through the Hausdorff distance against distortion
of the JPEG compression with the quality factor 10%, Additive white Gaussian
Noise (AWGN) of Signal-Noise Ratio (SNR) 10%, Rotation 10 ~ 15
o, a median filter
of window size 3x3, and low pass filter as mentioned in Table 4.1. Table 4.2
tabulates the quantitative deviation as the Hausdorff distance between the hash
values of the original and manipulated images for various attacks. The deviations are
less than 0.3 values except for large rotation (greater than 10o) and AWGN of SNR
(more than 10%).
47
Figure 4.2: Fingerprint images from FVC2004/DB1_A database
Table 4.1: Different Attacks used to assess the End Stopped Feature point
Attack
Parameters
Image Processing Operations
JPEG lossy compression
Additive white Gaussian Noise (AWGN)
Median filtering
Low pass filtering
Quality Factor =10
Standard deviation ๐ = 10
Window size 3x3
/
Geometric Distortion
Rotation
Degree 10 ~ 15
0
48
Table 4.2: Hausdorff Distance between features of original and Attacked Image
(For 32 Feature Points)
Attack 00_1 01_1 02_1 03_1
JPEG,QF=10 0.2524 0.2294 0.1205 0.1669
AWGN,๐ = 10 0.8522 0.3151 0.3546 0.2577
Rotation 10 0.1587 0.1196 0.1620 0.1371
Rotation 20 0.2356 0.1856 0.1884 0.1654
Rotation 30 0.2065 0.1935 0.1729 0.1811
Rotation 40 0.2743 0.2254 0.2803 0.1742
Rotation 50 0.2827 0.2221 0.2768 0.2183
Rotation 100 0.2698 0.3341 0.3204 0.2913
Rotation 150 0.3490 0.4066 0.3892 0.3926
Median Filter (3x3) 0.2092 0.0983 0.1116 0.1852
Low pass Filter 0.0649 0.0372 0.0435 0.0444
4.3 Speeded-Up Robust Features [SURF]
SURF method has have been used in general object recognition and other machine
vision applications for a number of years. The feature vectors are formed by means
of local patterns around key points detected using scaled up filter size. These
extracted feature vectors are established to be distinct and robust to noise and
geometric and photometric deformations of image [40]. The major steps for SURF
feature vectors are determined by key-point detectors and descriptors.
a) Key-Point Detectors
SURF detection based on Hessian Matric [89] leads to the use of integral images,
which drastically reduces the computation time. In addition, integral images fit in the
more general framework of boxlets [90]. The entry of an integral image ๐ผฮฃ ๐ at a
location ๐ = ๐ฅ, ๐ฆ represents the sum of all pixels in the input image ๐ช within a
rectangular region by the origin and ๐.
๐ผฮฃ ๐ = ๐ผ ๐, ๐ ๐ฆ๐ =0
๐ฅ๐=0 (4.10)
49
It can be said that Hessian-based detectors are more stable and repeatable than
their Harris-based counterparts. Given point ๐ = ๐ฅ, ๐ฆ in an image ๐ช, the Hessian
matrix ๐ป ๐, ๐ in ๐ at scale ๐ is defined as follows:
๐ป ๐, ๐ = ๐ฟ๐ฅ๐ฅ ๐, ๐ ๐ฟ๐ฅ๐ฆ ๐, ๐
๐ฟ๐ฅ๐ฆ ๐, ๐ ๐ฟ๐ฆ๐ฆ ๐, ๐ (4.11)
Where ๐ฟ๐ฅ๐ฅ ๐, ๐ is the convolution of the Gaussian second order derivative
๐2
๐๐ฅ 2 ๐(๐) with the image ๐ช in the point ๐, and similarly for ๐ฟ๐ฅ๐ฅ ๐, ๐ and ๐ฟ๐ฅ๐ฆ ๐, ๐ .
The second order Gaussian derivatives are approximated using box filters as
shown in Figure 4.3 and image convolutions with box filters are computed
rapidly using integral images. The determinant of Hessian matrix ฮH can be
reduced to
ฮH = ๐ท๐ฅ๐ฅ ๐ท๐ฆ๐ฆ โ (๐ค๐ท๐ฅ๐ฆ )2 (4.12)
The response for each spot can be determined by assigning ๐ค = 0.9 [15].
Furthermore, keypoints are localized in scale and image space by applying non-
maximum suppression in a 3x3x3 neighbourhood.
๐ฅ-direction ๐ฆ-direction ๐ฅ๐ฆ direction
Figure 4.3: Approximation of the second order Gaussian partial derivative.
(The grey regions are equal to zero)
50
b) SURF Descriptor
This stage consists of two sub steps. In the first sub step, a circular region is
constructed around the extracted key-points and the dominant orientation of the
circular region is computed using the Haar wavelets response in both ๐ฅ and ๐ฆ
directions. The resulting maximum Haar wavelet response is considered to be the
dominant orientation and is used to generate the key-point feature vector. The feature
vectors of a key-point are measured relative to the dominant orientation and thus, the
generated feature vectors are invariant to image rotations.
In the second sub step, a square region is constructed around each extracted key-
point and aligned along the dominant orientation. This square region is
partitioned into 16 sub-regions of size 4x4 and therefore, Haar wavelet responses
are computed for each sub-region. The sum of the wavelet responses, dx and dy,
for each sub-region are used as feature values. Further, absolute values ๐๐ฅ and
๐๐ฆ are summed up to obtain the polarity of the image intensity changes. Thus,
the descriptor vector, Des, of the sub-image is given as
Desโ ๐๐ฅ ๐๐ฆ ๐๐ฅ ๐๐ฆ (4.13)
The SURF descriptor vector of the keypoint is formed by concatenating descriptor
vectors of all sixteen 4x4sub-regions around a keypoint. In addition, it consists of 64
elements.
4.3.1 Experimental Results
SURF [15] focuses on the spatial distribution of gradient information within the local
point neighbourhood. SURF is a rapid, scale and rotational invariant detector and
descriptor. To verify the effectiveness of the SURF algorithm, the experiment
conducted on fingerprint images and a face image of the same subject for known
attacks of image processing and geometric operations as mentioned in Table 4.3. The
result focuses on matching the two images (i.e., fingerprint and face) on various
distortions with limited feature points. Figures 4.4 and 4.5 illustrate the feature
51
descriptors and matching of feature points (30 points) of the fingerprint image and
face for various distortion like rotations (5degree, 20degree and 180degree),
translation (25x25 window), histogram equalization, Median filter (5x5 window),
JPEG (10%) and Additive White Gaussian Noise with a SNR of 0.065%. It was
observed that descriptor vectors are matched between original and manipulated
images for various attacks.
Table 4.3: Different Attacks used to assess the SURF Feature point
Attack Parameters
Image Processing Operations
JPEG lossy compression
Additive white Gaussian Noise
Median Filter
Histogram equalization
Quality Factor =10
SNR of 0.065 %
Window size 5x5
/
Geometric Distortion
Rotation
Translation
Degree 00 ~ 180
0
Window size 25x25
52
(a) Feature Descriptor (b) Feature Points (379 Points)
(c) Feature Points (d) 5 degree Rotation ( e) 20 degree Rotation
(f) 180 degree Rotation (g) Translate (25x25) (h) Histeq
(i) Median Filter (5x5) j) JPEG (10%) (k) AWGN (0.065%)
Figure 4.4: Feature Descriptors and Matching of Feature Points (30Points) on
Fingerprint Images of the same subject for different attacks: (a) Feature Descriptor
(b) Feature Points (379 Points) (c) Feature Points (d) 5 degree Rotation e) 20
degree Rotation (f) 180 degree Rotation (g) Translate (25x25) (h) Histeq (i) Median
Filter (5x5) j) JPEG (10%) (k) AWGN (0.065%).
53
(a) Feature Descriptor (b) Feature Points (379 Points)
(c) Feature Points d) 5 degree Rotation e) 20 degree Rotation
(f) 180 degree Rotation (g) Translate (25x25) (h) Histeq
(i) Median Filter (5x5) j) JPEG (10%) (k) AWGN (0.065%)
Figure 4.5: Feature Descriptors and Matching of Feature Points (30 Points) on Lena
Images of the same subject for different attacks: (a) Feature Descriptor (b) Feature
Points (379 Points) (c) Feature Points (d) 5 degree Rotation e) 20 degree Rotation
(f) 180 degree Rotation (g) Translate (25x25) (h) Histeq (i) Median Filter (5x5) j)
JPEG (10%) (k) AWGN (0.065%).
54
4.4 Scale Invariant Feature Transform [SIFT]
Local features, such as corners, blobs and regions, have been widely used for object
detection, recognition and retrieval purposes in computer vision. The intrinsic
advantages of these local features are their invariance under geometric operations.
Among various local feature detectors and descriptors, SIFT was shown to provide a
relatively optimal trade-off between robustness, unique and efficiency.
SIFT [51] mainly consists of the following steps to generate the set of an image
feature: Scale-space extrema detection, keypoint localization, orientation assignment
and keypoint descriptor.
a) Scale-Invariant Points Detection and Localization
Scale invariant local points are detected by searching for local extrema in the series
of difference-of-Gaussian (DoG) image. The DoG is constructed by the convolution
of a variable scale Gaussian function G(x, y, ฯ), with an image I(x, y) in the scale
space of an image L(x, y, ฯ).
๐ฟ ๐ฅ, ๐ฆ, ๐ = ๐บ ๐ฅ, ๐ฆ, ๐ โ ๐ผ(๐ฅ, ๐ฆ) (4.13)
To efficiently detect stable keypoint locations in the scale space, using scale-space
extrema in the DoG convolved with the image, ๐ท ๐ฅ, ๐ฆ, ๐ with a nearby scale
๐ฯ and ฯ as
๐ท ๐ฅ, ๐ฆ, ๐ = ๐ฟ ๐ฅ, ๐ฆ, ๐ฯ โ ๐ฟ ๐ฅ, ๐ฆ, ฯ
= ๐บ ๐ฅ, ๐ฆ, ๐ฯ โ ๐บ ๐ฅ, ๐ฆ, ฯ โ ๐ผ ๐ฅ, ๐ฆ (4.15)
Essentially, the DoG detector could be attributed to the detector for blob structures in
the image content, as it provides a close approximation of the scale-normalized
Laplacian of Gaussian.
๐บ ๐ฅ, ๐ฆ, ๐ฯ โ ๐บ ๐ฅ, ๐ฆ, ฯ โ (๐ โ 1)๐2โ2๐บ (4.16)
Substituting (4.16) into (4.15) and using a property of convolution to obtain
55
๐ท ๐ฅ, ๐ฆ, ฯ โ ๐ โ 1 ๐2โ2๐บ โ ๐ผ ๐ฅ, ๐ฆ
= ๐ โ 1 ๐2โ2 โ โ2๐ผ ๐ฅ, ๐ฆ (4.17)
where the Laplacian operator โ2๐ผ ๐ฅ, ๐ฆ is used to detect edges and corners in the
images. Equation (4.17) is DoG, is approximation of Laplacian of Gaussian and
provides greater robustness against geometric transform of images compared with
other gradient based feature points detectors such as Harris and Hessian, etc.
b) Orientation Assignment and Descriptor Generation:
The orientation of each key-point is determined by the peak of the orientation
histogram formed by the gradient orientations within its neighbourhood. Based on
the position, scale and orientation of each key-point, the corresponding descriptor
with 128 dimensions based on gradient histogram within its 16x16 local
neighbourhood is generated.
4.5 SIFT-Harris
DoG detector of SIFT provides a satisfying performance under geometric
transforms; however, its robustness against attacks, like additive noise and blurring,
is poor [16]. To extract robust local features, it is desirable to select the most stable
key-points under various distortions and attacks. The Harris corner [91] could
provide a stable detection performance with high repeatability and localization
accuracy under various distortion and geometric transformations. Therefore, the
Harris criterion is incorporated to select the most stable SIFT keypoints.
The Harris detector / criterion is based on the auto correlation matrix, which
represents the gradient distribution within a local region of the selected point. The
autocorrelation matrix M for image ๐ผ ๐ฅ, ๐ฆ is represented as
๐ = ๐ค(๐ฅ, ๐ฆ)๐ฅ ,๐ฆ ๐ผ๐ฅ
2 ๐ผ๐ฅ๐ผ๐ฆ
๐ผ๐ฅ๐ผ๐ฆ ๐ผ๐ฆ2 (4.18)
Where ๐ค(๐ฅ, ๐ฆ) is a window to determine the accumulated region, and ๐ผ๐ฅ and ๐ผ๐ฆ are
the image gradients in ๐ฅ and ๐ฆ axis, respectively. Furthermore, there is alternative
criterion to evaluate the corner points as;
56
H=๐1๐2 โ ๐พ(๐1 + ๐2)2 = det ๐ โ ๐พ(๐ก๐๐๐๐2 ๐ (4.19)
where ๐1,๐2 are eigenvalues of autocorrelation matrix ๐ and ๐พ is a coefficient
ranging from 0.04 - 0.15 empirically. In this research work, ๐พ = 0.06 [16].
Given a set of SIFT points = ๐๐ ๐ฅ, ๐ฆ, ๐, ๐ ๐=1๐ , where ๐ฅ and ๐ฆ are coordinates and
๐, ๐ are scale and orientation parameters, respectively. ๐ป๐๐ (๐ฅ, ๐ฆ) is the Harris
response where ๐ is the standard deviation of the Gaussian kernel window used to
compute the auto-correlation matrix M and set the threshold ๐ to select the robust
SIFT point as
๐ = ๐ผ
๐ ๐ป๐
๐ ๐ฅ, ๐ฆ
๐
๐=1
4.20
where ฮฑ is an adjustable parameter to control the robust points selection and
empirically ฮฑ โ [0.1, 0.5]. Therefore, to control the robustness 0.5 is chosen as the
default value.
4.5.1 Experimental Results
SIFT feature key points consist of local maxima or minima together with the gradient
histogram. During matching, all the SIFT key points of the two images are
compared. Intuitively, the SIFT algorithm is able to localize objects in an image, it
can also help to determine whether two images contain identical content. Based on
such consideration, the algorithms for SIFT and SIFT-Harris are tested on a natural
image of size 256x256, between the original image and known distortion like
Gaussian Noise ( var =0.005), Gaussian Blurr (var =0.5, 5x5 window), JPEG (QF
=10%) as illustrated in Table 4.4.
57
Table 4.4 Comparison of keypoints Detectors on original and distorted Image
Attacks
SIFT
SIFT-Harris
Original
Image
Keypoints
Distoted
Image
Keypoints
Original
Image
Keypoints
Distoted
Image
Keypoints
Gaussian Noise
(var =0.005)
306
370
87
117
Gaussian Blurr
(var =0.5, 5x5)
353 117
JPEG (QF =10 %) 460 148
Furthermore, an investigation is performed on the benefits of robust keypoints
against content preserving manipulations for the purpose of content identification. In
this analysis Hausdorff distance is used to measure the similarity between the
coordinates of the two sets of features detected in the original and distorted images.
Alongside this, Hausdorff is used to compare the state of the art feature detector
SIFT-harris with end stopped wavelet detector in terms of robustness against content
preserving manipulation, as shown in Table 4.5. The feature vector are coordinates of
the top 20 detected stable keypoints. It is observed that the average Hausdorff
distance for a keypoint detected by the SIFT-Harris detector are smaller than an end
stopped wavelet.
Table 4.5 Average Hausdorff Distance between the coordinate for the top 20
keypoints detected in the original and manipulated copies using the SIFT-harris and
End Stopped Detector.
Manipulation SIFT Harris End Stopped
Gaussian Noise (Var =0.005) 0.0039 0.0567
Salt and pepper Noise (Var =0.01) 0.0008 0.0902
Speckle Noise (Var =0.01) 0.0013 0.0066
JPEG, QF =10 0.0021 0.0131
58
4.6 SIFT-Wavelet
Content Based Image Retrieval (CBIR) plays significant role in image processing
and its mainly focus on describing the bottom information of images, such as color,
texture, etc., These methods have some achievements but they have more difficulties
to describe image scaling, rotation movement, affine and other features in detail.
One of the appealing methods is to collect the salient points (feature points) using
low-level characteristics such as Harris detector and SIFT. Salient points can also be
detected on the wavelet domain. Wavelet based technique is basically used to
provide more security and reliability of image. It decomposes an image into various
resolutions which provide approximate and detail coefficients of image, which is
then further processed for feature extraction and matching. In this proposed research
we combined SIFT feature with efficient wavelet-based salient points to generate
robust SIFT - wavelet feature that provides sufficient invariance to common image
manipulations. The detail literature and the proposed block diagram are explained
in section 5.4.2
4.7 Overview of Feature Detector
An invariant feature is an image pattern which differs from its immediate
neighbourhood and is usually associated with a change in an image property, such as
intensity, colour and texture. A good set of features holds the following properties:
Repeatability: in which the feature extraction process should be repeatable and
precise, so that the same features are extracted on two images showing the
same object.
Distinctiveness: is the intensity pattern underlying the feature variations,
which can be distinguished and matched.
Locality: allows the features invariant to reduce the probability of occlusion
and model approximations of the geometric and photometric deformations
between two images at different viewing conditions.
59
Quantity: the number of detected features should be sufficiently large, so that
a reasonable number of features are detected, even on small objects.
Accuracy: the detected features should be accurately localized, both in image
location, with respect to scale and possibly shape.
Efficiency: the detection of features in a new image should allow for time
critical applications.
Table 4.6 [49] provides an overview of the feature detector and has been grouped
according to their invariance: rotation, similarity, affine and perspective. The Harris
detector has rotational invariant features with the highest repeatability and
localization accuracy. The Hessian detector locates blobs which are not as well
localized and requires second-order derivatives to be computed. The SUSAN
detector avoids computation of derivatives and is known for its efficiency; however,
the absence of smoothing makes it more susceptible to noise.
In scale, invariant feature group Harris-Laplace has been shown to have high
repeatability and localization accuracy inherited from the Harris detector. Hessian-
Laplace is more robust than its single scale nature, which is due to blob-like
structures that are better localized in scale than corners. In this invariant group, DoG
and SURF detectors were designed for efficiency. The DoG detector performs
extremely well in matching and image retrieval due to superior balance between
spatial localization and scale estimation accuracy.
The affine invariant Harris and Hessian follows the invariance properties. In
addition, the salient regions require computing of a histogram and its entropy for
each regions candidate in the scale or affine space, which results in large
computational costs. Thus, the edge based regions focus on corners formed by edge
junctions and provides good localization accuracy and repeatability with a fewer
number of detected features.
Intensity based regions use a heuristic method and find similar regions to the
Maximally Stable Extremal Regions (MSER). Superpixels are typically based on
segmentation methods, which are computationally expensive like normalized cuts. In
60
contrast to the superpixels, the MSER selects only the most stable regions, which
results in high repeatability.
Binary Robust Independent Elementary Features (BRIEF) descriptors are relies on a
relatively small number of intensity difference tests to represent an image patch as a
binary string. BRIEF construction and matching for this descriptor much fast to
yield higher recognition rates. The descriptor similarity can be evaluated using the
Hamming distance, which is very efficient to compute, instead of the L2 norm as is
usually done.
Binary Robust Invariant Scalable Keypoints (BRISK) tackles the classic Computer
Vision problem of detecting, describing and matching image keypoints for cases
without sufficient a priori knowledge on the scene and camera poses. BRISK relies
on circular sampling pattern from which it computes brightness comparisons to form
a binary descriptor string and offers the quality of high-end features in time
demanding applications.
Fast Retina keypoint (FREAK) is a fast, compact and robust keypoint descriptor. A
cascade of binary strings is computed by efficiently comparing pairs of image
intensities over a retinal sampling pattern. FREAKs are general faster to compute
with lower memory load and also more robust than SIFT, SURF or BRISK.
61
Table 4.6: Overview of Feature Detector [50]
Feature
Extractor Corner Blob Region
Rotation
Invariant
Scale
Invariant
Affine
Invariant Repeatability
Localization
Accuracy Robustness Efficiency
Harris +++ +++ +++ +
Hessian ++ ++ ++ +
SUSAN + ++ ++ +++
Harris-Laplace +++ +++ ++ +
Hessian-Laplace +++ +++ +++ +
DoG ++ ++ ++ ++
SURF ++ ++ ++ +++
Harris-Affine +++ +++ + +
Hessian Affine +++ +++ +++ +
Salient Regions + + ++ +
Edge-Based +++ +++ + +
MSER +++ +++ ++ +++
Intensity Based ++ ++ ++ ++
Superpixels + + + +
BRIEF ++ ++ ++ +++
BRISK ++ ++ ++ +++
FREAK ++ ++ +++ +++
62
4.8 Summary
In this chapter, the most recent state-of-the-art feature extraction techniques: end-
stopped wavelets, SURF, SIFT, and SIFT-Harris are investigated for their perceptual
robustness against various content-preserving manipulations. Based on geometric
invariance of the SIFT keypoints, Harris criterion are incorporated to select the most
stable feature points under the addition of noise and blurring. The performance of
this feature detector is evaluated through Hausdorff distance to measure the
similarity between the feature vector of the original and distorted images. The feature
vectors are coordinates of the top 20 detected stable keypoints. The average
Hausdorff distance of the keypoint detected by the SIFT-Harris detector are smaller
than end stopped wavelets. Therefore SIFT-Harris is relatively more stable for
geometric operations, especially for blurring operations and additive noise attacks.
This chapter also discusses the overview of feature detectors by highlighting their
respective strengths and weaknesses. In next chapter, we discuss implementation of
robust minutiae based fingerprint image hashing. The fingerprint minutiae extraction
method is combined individually with the SIFT-Harris, SIFT-Wavelet and the SIFT
method, to generate robust features. The idea is to incorporate the orientation and
descriptor in the minutiae of the fingerprint image. Shape context based hashing is
used for fingerprint identification.
.
63
Chapter 5
Minutiae Based fingerprint Image Hashing
Due to recent developments in technologies, image data in digital format is used
extensively in the world of the internet and ready to be accessed anytime, anywhere.
The amount of image data information via digital devices has grown exponentially.
Conversely, the digital nature of information allows individuals to manipulate and
duplicate data easily without any change in quality. The power of the image
manipulation software has made it possible to effortlessly modify digital multimedia
data. Under this circumstance, integrity verification has become an important issue in
the digital world. In the area of multimedia security, two types of approaches have
been proposed to maintain the confidentiality of image data: watermarking and
perceptual hashing. Watermarking is the ability to detect changes in the host image
data, which can provide some form of guarantee that the image data has not been
tampered with and has originated from the right source. Watermarking can be used in
copyright verification or in content authentication for digital images. Furthermore,
data embedding inevitability causes a slight distortion in the host image data.
In conjunction with watermarking, perceptual hashing is conventionally used for
image content authentication. The main advantage of a perceptual hashing scheme is
that the image data is not altered and not degraded visually. Perceptual image
hashing is different from cryptographic hashing in that cryptographic hashes are
extremely sensitive to single-bit changes of input data that will change the output
dramatically.
Recently, perceptual image hashing has received considerable attention from many
researchers. Most countermeasures proposed in the literature, generally focus on
feature extraction to acquire robust feature to authenticate the image. In this research,
we introduced a robust minutiae based fingerprint image hashing by combining state
of the art feature points with the minutiae of fingerprint images, which are discussed
in the following section 5.5.
64
5.1 Design criteria for image hashing
There are three important design criteria for an image hash function: robustness,
fragility and unpredictability. Let I denote the reference image (e.g., all natural
images of a particular size). Likewise, I' denote test image (an image visually
identical or perceptually distinct). Assuming hash function H (.), the system produces
two hash values h = H(I) and h' = H(I') by using a key K.
๐ = ๐ป(๐ผ; ๐พ)
๐โฒ = ๐ป(๐ผโฒ;๐พ) (5.1)
The following requirements are considered:
(i) Perceptual robustness:
(H(I,K) โ H(I',K)),then dist(h, h') <Th (5.2)
(ii) Distinct visually:
(H(I,K) โ H(I',K)),then dist(h, h') โฅ Th (5.3)
(iii) Unpredictability of the hash:
H(I); fh(1) = fh(0) = 0.5 (5.4)
Where,
fh(x) is the probability mass function for h.
In effect, (i) ensures the capability of the visually similar images distance between
two hashes h and h' should be less than a threshold Th. Conversely, (ii) has the
capability to differentiate between two images, which are visually distinct. The first
requirement suggests that the distance between two hashes h and h' should be larger
than a threshold Th. (iii) guarantee the unpredictability of the hash values, in addition
to the key K functions (pseudo-random number generator) used at final stages of the
hash function.
65
Figure 5.1: Design requirements for fingerprint image hashing
5.2 Perceptual Image Hashing Schemes
Perceptual image hashing has been introduced to generate a robust, unique, compact
and secure feature and moreover, its hash values of the image. Based on the
characteristics of images, the extracted features are unique and distinctive for content
identification. The image hashes are perceptually similar to the content preserving
operations as long as two images are similar to the Human Visual System (HVS).
Similarly, a very different hash value for a perceptually different image is shown in
Figure 5.1. Hence, feature extraction is a key step in the image hashing technique.
The image hashing schemes are categorised into four types: Statistic-based schemes,
Relation-based schemes, Coarse-representation-based schemes and Low-level
feature-based schemes.
66
Statistic-based schemes [43], [46], [47], [92]: This group of schemes extracts
image features by calculating the image statistics in the spatial domain, such
as the mean, variance, higher moments of image blocks and histogram.
Relation-based schemes [90], [94]: This category involves approaches to
extract image features by making use of some invariant relationships of the
coefficients of the discrete cosine transform (DCT) or wavelet transform
(DWT).
Coarse-representation-based schemes [41], [48], [95]: In this category, the
perceptual hash is calculated by making use of coarse information with
regards to the whole image, for instance the spatial distribution of significant
wavelet coefficients, the low-frequency coefficients of Fourier transform, and
so on.
Low level feature-based schemes[14], [96]: The image features are extracted
by detecting the salient image feature points. These methods first perform the
DCT or DWT transform on the original image, and subsequently, makes use
of the coefficients to generate a final hash value. However, the perceptual
hash value is very sensitive to global as well as local distortions that do not
cause perceptually significant image changes.
5.2.1 Perceptual Image Hashing Framework
Transformation QuantizationFeature Extraction
Crypto-
Compression
Input imageTransformed
image
Continuous
intermediate
Hash vector
Intermediate
Binary hash
vector
Final Perceptual
hash
T( ) FX( ) Q( )
En( )
Figure 5.2: Pipeline stages for the perceptual hashing system
67
A perceptual hashing system consists of four pipeline stages: transform stage, feature
extraction stage, quantization stage, and compression and crypto-compression stage,
as illustrated in Figure 5.2. In the transformation stage, the input image undergoes
special and /or frequency transformation to make all the extracted features depend on
the image pixels or image frequency coefficients. In the feature extraction stage, the
perceptual image hashing system extracts the image feature from the input image to
generate the continuous hash vector. They, the continuous perceptual hash vector are
quantized into the discrete hash vector in the quantization stage. In the third stage the
discrete hash vector is converted into a binary perceptual hash string. Finally, the
binary perceptual hash string is compressed and encrypted into a short and final
perceptual hash in the crypto-compression stage.
5.3 Content-Based Image Identification
The concept behind image authentication is to extract the image characteristics
(features) of the human perception for the authentication process. Typically, some
applications will be performed by considering intentional (image processing, such as
filtering, compression, cropping, resizing etc) and nonโintentional (noise, channel
errors) distortion to the images. The context-preserving manipulations only change
the pixel values, which results in different levels of visual distortion in the image, but
the contexts of the image, which carries the same visual message to the observer, are
preserved. In contrast, malicious/content-changing manipulations consist of changing
the content of the original image to a new one, which transmits a different visual
message to the observer. One typical example of malicious modification is replacing
some parts of the image with different contents. Perceptual hashes are expected to be
able to survive acceptable content-preserving manipulations and
reject malicious manipulations. Classification of content-preserving and content-
changing manipulations is provided in Table 5.1.
68
Table 5.1: Content-Preserving and Content-Changing Manipulation
Content โPreserving Manipulation Content โChanging Manipulation
Transmission errors
Noise
Compression and quantization
Resolution reduction
Scaling
Rotation
Cropping
๐พ Distortion
Colour conversions
Contract adjustment, changes of brightness, hue and saturation
Removal of image objects
Moving of image elements or
changing their positions
Adding new objects
Changes of image characteristics:
colour, textures.
Changes to the image
background: day time or location
Changes of light conditions:
shadow manipulation etc.
5.4 Proposed Feature Extraction Techniques
The state of the art feature extraction techniques and the properties of feature
detector are briefly discussed in chapter 4.
This research focuses on improving the security of fingerprint templates and accurate
comparison of the fingerprint content. The generation of fingerprint templates, which
in turn are used to compare fingerprint content (or their perceptual hashes) mostly
rely on feature extraction techniques, such as SIFT, SIFT-Harris or Fingerprint
Minutiae. However, a combination of the two (e.g., SIFT-Minutiae) has not been
previously studied in the literature.
Firstly, the SIFT-Harris method is combined with the Fingerprint Minutiae extraction
technique to determine the most prominent fingerprint features. These features are
post-processed into perceptual hashes using shape content based perceptual hashing
to plot the accuracy of fingerprint comparison using Receiver Operating
Characteristic (ROC) curves.
69
Secondly, the SIFT-Wavelet is combined with the Fingerprint Minutiae extraction
technique to determine the most prominent fingerprint features. These features are
also post-processed using shape content based perceptual hashing techniques to plot
the accuracy of fingerprint comparison using ROC curves.
Thirdly, the SIFT is combined with the Fingerprint Minutiae extraction method and
post-processed using shape content based procedures to plot the accuracy.
5.4.1 SIFT-Harris-Minutiae Feature for Fingerprint Image Hashing
To design robust fingerprint hashing against various distortions, robust feature
extraction is the most important step. We propose a method for extracting the robust
minutiae of the fingerprint images, by combining the SIFT-Harris feature points with
the minutiae of the fingerprint image, as demonstrated in Figure 5.3. Based on the
position, scale and orientation of each keypoint in the SIFT-Harris, the corresponding
descriptor with 128 dimensions based on the gradient histogram within its 16x16
local neighborhood is generated. Alongside the SIFT-Harris feature points (SHFP),
minutiae (MP) are extracted with four tuple, such as MP={x,y, ๐, t}, where, (x,y) is
the coordinate, ๐ is angle, and t is the type of minutiae (termination or bifurcation)
respectively.
Fingerprint
Minutiae
Extraction
SIFT -Harris
Feature
Extraction
Fingerprint Image
Sets
Feature Point
Matching
Feature
Compression
Feature
Hashes
Secret Key
Query
Hashes
DecisionVerification/
Identification
Robust Feature
Figure 5.3: Proposed Robust SIFT-Harris- Minutiae based Fingerprint Image
Hashing
70
The absolute radial distance between the position of the SIFT-Harris feature and the
coordinates of the minutiae are computed. Most robust minutiae are detected if the
relative difference of pixel values is within the threshold of ten pixel value [5]. Along
with the detected point the corresponding orientation and descriptor are also
identified. Hence, the proposed robust minutiae are represented by coordinates,
orientation and descriptor respectively. This robust minutia of the fingerprint image
is to be hashed, and the fingerprint identification is to be performed using hashed
robust minutiae. Consequently, the detailed hashing technique is explained in section
5.5.
The robustness of the image hashing arises from robust feature extraction and the
compression, which mainly contributes to the compactness of the final hash. To
increase the security of a traditional hash function and prevent unauthorized access, a
secret key is incorporated in either the feature extraction, compression or both to
make the hashes unpredictable.
Most of the hashing algorithms incorporate a pseudorandomization relying on a
secret key. Such a key is incorporated into the compression step [15] to further
enhance the security, as indicated by the dashed line in Figure 5.3. The key is owned
by the owner, and the hash generation is a pseudorandom process rather than a
completely random one for fingerprint identification. The incoming query hash
corresponding to the query image is compared with the hashes in the database.
5.4.2 SIFT-Wavelet -Minutiae Feature for Fingerprint Image Hashing
Recently, the wavelet based salient points detector and a combination of wavelet-
SIFT features has been successfully used in several image recognition systems. Lin
et al [97] presented an efficient salient-region extraction algorithm based on the
significance of accumulated wavelet coefficients. This algorithm is robust for image
processing operations like compression, filtering and geometric distortions. Lim [98],
Sebe et al [99] explained the comparative analysis of scale-invariant feature
extraction using different wavelet bases and highlight the advantages, whereas
Loupias et al [100] recommended a salient point detector based on wavelet
transform to detect global and local variations. In addition, Omidyeganeh et al [101]
71
introduced a robust face recognition system based on wavelet transform of the Scale
Invariant Feature Transform (SIFT) features, and moreover, these combinations
provide important performance efficiency in face recognition systems.
Recently, Kumar et al [102] suggested a new technique for robust colour image
matching based on the combination of wavelet-colour SIFT features. Moreover,
Wang et al [103] presented a parameter adjusted Gaussian Mixture Models using a
salient feature patched to recognise an object in Caltech image database. These
methods combine an effective combination of wavelet-based features and SIFT
features. Thus, the extracted features are suitable for properties that are invariant to
translation, scaling, rotation and illumination changes.
Halawani et al [104] evaluated and compared the non-linear kernel function around
salient points with Monte-Carlo[105] function for the purpose of invariant content
based image retrieval. Jian et al [106] proposed a wavelet based salient point detector
to extract the visually meaningful region in an image and an annular segmentation
algorithm based on salient region distribution is designed. Furthermore, Imamoglu et
al., [107] introduced a novel saliency detection model by utilizing low-level features
obtained from the wavelet transform domain.
Salient point detection in images is very useful for image processing application like
image compression, object detection and object recognition. Tsai et al [108]
suggested a hierarchical selection algorithm for selecting the most salient point to
satisfy the representation of an image and to make an effective image retrieval
system.
Though SIFT algorithms locate the salient points in a given image which are scale
invariant there is a need to improve the robustness of those points. The most
appropriate method for doing this is to prune the locations obtained by the SIFT to
retain the most robust points, which remain unchanged for different types of attacks.
Consequently, it is demonstrated [16] that the corner points in the SIFT locations are
highly robust and retain the stable points.
72
A) Wavelet Transform
Wavelet [109] plays a significant role in many image processing applications. The
computation of the wavelet transform of a 2-D image involves recursive filtering and
sub-sampling. This operation results in four decomposed subband images referred to
as low-low (LL, produces a approximation of the image), low-high (LH, containing
horizontal information at a high frequency), high-low (HL, containing vertical
information at a high frequency), and high-high (HH, containing diagonal
information at a high frequency). Moreover, the wavelet transform can recursively
decompose the LL band. The two level wavelet decomposition results, LH1, HL1,
HH1, LH2, HL2, HH2 and an additional approximation image LL2 are revealed in
Figure 5.4. In this particular research, we use the Haar wavelet [110], which is a
simple orthogonal, compactly supported wavelet, which leads to a complete and non-
redundant representation of the image.
LL2 LH2
LH1 HL2 HH2
HL1 HH1
Figure 5.4: Second Level Wavelet Transform
B) Proposed SIFT-Wavelet -Minutiae
In our work we intend to use a wavelet image analysis tool to determine the robust
points in the SIFT. This is completed by the motivation that four categories of
coefficients; approximation coefficients, horizontal, vertical and diagonal
coefficients can be utilized to localize the salient points. This method can be
combined with the SIFT to select the robust points from the SIFT coefficients.
To explain the proposed method of selecting robust points from the SIFT output, let
us consider that the selected images are decomposed by the L levels and represent
73
the coefficients obtained for every level, as ak, hk, vk, dk. Moreover, the relationship
between the coefficients at every level can be illustrated as in Figure 5.5.
Figure 5.5: Image Decompose Level
It can be perceived that if the input image is of dimension N x N then at level 0 each
of the coefficients will be of dimension N/2 x N/2. Furthermore, the computation 2 x
2 pixel group of the image contributes for a pixel in the set {ak, hk, vk, dk} k=0. Based
on image decomposition (Figure 5.5) it can be observed that when we compute the
next level coefficients the results are contributed by a 4x4 pixel group of the image.
When the decomposition is done progressively it can be shown that at kth level the
contribution for a pixel of coefficients would be from a b x b pixel group from the
input image, whereas b=2(k+1).
We propose to use the coefficient value obtained at a location at the kth level to
determine the characteristics of the b x b pixel group of the original input image. If
all the three coefficients hk, vk, dk are high at kth level in a given location then we
assume that the corresponding b x b pixel group of original image is salient. The
simple logic we applied is that a SIFT point in a salient bxb matrix of the original
image will be considered salient and retained. The high energy in horizontal, vertical
and diagonal directions imply that the point is a corner point and will be robust to
attacks.
The SIFT-Wavelet is combined with the Fingerprint Minutiae extraction technique to
determine the most prominent fingerprint features, as shown in Figure 5.6. The
algorithm is similar to the SIFT-Harris-Minutiae as discussed in section 5.4.1.
ao
ho
vo
do
a1
h1
v1
d1
a2
h2
v2
d2
Level-0 Level-1 Level-2
74
The proposed robust minutiae are represented by coordinates, orientation and
descriptor, respectively. The robust minutiae of the fingerprint image are also post-
processed using shape content based perceptual hashing techniques to plot the
accuracy of fingerprint comparison using ROC curves, as explained in section 5.6.
Fingerprint
Minutiae
Extraction
SIFT-
Wavelets
Feature
Extraction
Fingerprint Image
Sets
Feature Point
Matching
Feature
Compression
Feature
Hashes
Secret Key
Query
Hashes
DecisionVerification/
Identification
Robust Feature
Figure 5.6: Proposed Robust SIFT-Wavelet- Minutiae Fingerprint Image Hashing
5.4.3 SIFT-Minutiae Feature for Fingerprint Image Hashing
Together with the SIFT-Harris-Minutiae and the SIFT-Wavelet-Minutiae, the
performance of the SIFT-Minutiae combination was demonstrated. Without any
thresholding to the SIFT keypoints the minutiae of the fingerprint image was
combined to extract the most robust minutiae from the fingerprint images. The
position, scale, orientation and descriptor of each keypoint in the SIFT and minutiae
coordinate, angle and type are computed. The absolute radial distance between the
positioned SIFT points and the coordinates of the minutiae are calculated. Based on
the relative difference of the pixel the most robust minutiae are detected along with
the orientation and descriptor. In addition, the SIFT-Minutiae combination is
represented by the coordinates, orientation and descriptor respectively. This robust
minutia from the fingerprint image is post-processed using shape content based
procedures to plot the accuracy, as described in section 5.5.
75
5.5 Image Hashing Based On Shape Contexts
Roy et al [111] presented a technique to encode the geometric relationship between
the SIFT points into a short vector, but the robustness is limited to attacks like
rotation, cropping and compression. In [112], [113] the authors use the SIFT local
feature points to detect image copies or near-duplicate copies by matching the high-
dimensional local feature descriptors of keypoints. However, this is not possible in
image hashing, where, the robust features are compressed into compact hash and
match the hashes during the detection stage.
Lv et al [16] recommend using the shape contexts, which is a promising method to
measure shape similarity for object recognition, so as to generate image hashes based
on the robust local feature points that have been detected. The distribution of local
feature points is composed of the content structure of images and considers this
geometric structure as an abtract object. Furthermore, the use of a descriptor
represents this structure as an unique signature. In the following sections, the basic
concept of shape contexts and shape contexts based image hashings are described in
detail.
5.5.1 Shape Contexts
Given a set of points P= ๐๐ ๐=1๐ , which are sampled from the contour of an object, the
shape context of point ๐๐with respect to the reference point ๐๐ is defined in [114] as:
hi(k) = #{pi โ pc : (pi โ pc ) โ bin(k)} (5.5)
where pi โ ๐ and ๐๐๐(๐) are uniform in log-polar coordinates as shown in Figure 5.7
with the centre located at pc . The shape context of each point is a coarse histogram,
which represents the relative positions of other points to the reference point. It has
been identified that this descriptor is highly robust to shape deformation and offers a
globally discriminative characterization, which is effective in solving shape matching
and transforming model estimation problems.
76
Figure 5.7: Diagram of the original shape contexts and the proposed shape contexts
hashing: RSCH.ASCH. (a) Original Shape Contexts. (b) Radial Shape Contexts
hashing (c) Angular shape Contexts hashing [16].
5.5.2 Shape Context Based Hashing
In our proposed approach, shape context based hashing [16] is used for fingerprint
authentication and provides an excellent description of the geometric structure of a
shape. We can embed the geometric distribution of robust feature points, as well as
their descriptors into shape contexts to generate a compact image hash. The original
shape context was designed to be computed for each point sampled from the object
contours, which means that for N local feature points we have N shape contexts. In
addition, it provides a rich descriptor to represent the shapes, although it has to be
compressed to be used for hashing directly.
It can be observed in the image content authentication that all perceptually
insignificant distortions and malicious manipulations on the image content would not
lead to viewpoint changes. In addition, the centre of an image is generally preserved
and relatively stable under certain geometric attacks. This encourages Lv et al [16] to
generate shape contexts with the reference point in the centre and obtain a compact
signature for the image. Another reason for avoiding computing shape context for
each local feature point in the hashing is that the detection of keypoints cannot
guarantee to yield exactly the same feature points when the image is under different
attacks and manipulations. As a trade-off, Radial Shape Context Hashing (RSCH)
and Angular Shape Context Hashing (ASCH) are used to generate hashes using
shape contexts with respect to the central reference point.
77
Given a set of robust feature points ๐ = ๐๐ ๐ฅ, ๐ฆ ๐=1๐ and their corresponding
descriptors ๐ท = ๐๐๐ (x, y)}๐=1๐ , the basic steps of RSCH and ASCH are as follows :
A. Radial Shape Context Hashing (RSCH)
a. Given the coordinates of the central point ๐ถ = (๐ฅ๐ ,๐ฆ๐) and the required length of
the hash L, construct bins ๐ต = ๐ ๐ ๐=1๐ of shape contexts with incremental
๐ =max (๐ฅ๐ ,๐ฆ๐)/L in the radial direction of the polar coordinates.
๐(๐) = {๐๐ โ ๐: ๐ โ 1 ๐ โค ๐๐ โ ๐ถ โค ๐๐} (5.6)
Where ๐๐ โ ๐ถ is the relative distance between pi and the central point ๐ถ.
b. Generate pseudorandom weights โ๐ ๐=1๐ฟ from the normal distribution N (๐ข, ๐2)
using a secret key. Each โ๐ is a random vector with 128 dimensions to be
consistent with the dimensions of the SIFT descriptors.
c. Let ๐ป = ๐๐ ๐=1๐ฟ be the hash vector and thus, we have each component ๐๐ as
๐๐ = ๐ค ๐ฟโ๐๐๐
2๐ โ๐ , ๐๐๐
๐๐โ๐ ๐
(5.7)
Where โ๐๐๐ = (๐๐๐ โ ๐๐) โ (0,2๐) is the relative difference in orientations
between ๐๐ and the central point๐ถ. The weight w [๐ฟโ๐๐๐ / 2๐] โ ๐ = ๐ค๐ ๐=1๐ฟ , is
the set of random weights generated from uniform distribution U(0.5,1). This is
to differentiate between the points located at different orientations of the same
hash bin ๐(๐) along the radial direction.
78
B. Angular Shape Context Hashing (ASCH)
a. Given the coordinates of the central point ๐ถ=(๐ฅ๐ ,๐ฆ๐) and the required length of
the hash L, construct bins ๐ต = ๐ ๐ ๐=1๐ of shape contexts with incremental
๐ = 2๐/L in the angular direction of the polar coordinates .
๐(๐) = {๐๐ โ ๐: ๐ โ 1 ๐ โค (๐๐๐ โ ๐๐) โค ๐๐} (5.8)
Where (๐๐๐ โ ๐๐) =โ๐๐๐ โ 0,2๐ .
b. Generate pseudorandom weights โ๐ ๐=1๐ฟ from the normal distribution N(๐ข, ๐2)
using a secret key. Each โ๐ is a random vector with 128 dimensions to be
consistent with the dimension of the SIFT descriptors.
c. Let ๐ป = ๐๐ ๐=1๐ฟ be the hash vector, we have each component ๐๐ as
๐๐ = ๐ค ๐ฟ ๐ ๐โ๐ถ
๐ถ โ๐ , ๐๐๐
๐๐โ๐ ๐
(5.9)
Where ๐๐ โ ๐ถ is the same as referred to in section 5.5.2A and ๐ถ = ๐ฅ๐2 + ๐ฆ๐
2 is
the normalization factor. The weight ๐ค[๐ฟ ๐๐โ๐ถ / ๐ถ ] โ ๐ = ๐ค๐ ๐=1๐ฟ is the set of
random weights generated from the uniform distribution U(0.5,1). This is to
differentiate between the points located at different orientations on the same hash bin
๐(๐) along the angular direction.
Estimation of Central Orientation ๐C : The central orientation ๐C is significantly
important for both the ASCH and RSCH and moreover, is required as a reference
direction to calculate the relative difference in orientation between the local feature
point ๐๐ and the central point ๐ถ. However, estimating based on local gradient
distribution is not reliable due to different image processing attacks. Alternatively,
Radon transform is used to estimate an accurate reference orientation for central
point ๐ถ.
Radon transform is the integral transform consisting of the integral of a function over
straight lines. Given a 2-D function ๐ (๐ฅ, ๐ฆ) and line ๐ with orientation ๐ as shown in
figure5.8
79
f(x,y)
y
x
p
q
Pr
ฮธ
R(p.ฮธ)
Figure 5.8 Radon transform ๐ (p,๐) of a 2-D function ๐ (๐ฅ, ๐ฆ)
The Radon transform of ๐ (๐ฅ, ๐ฆ) is the integral of the orthogonal projection to line ๐.
๐ ๐ (๐,๐)=โซ ๐ ๐ฅ, ๐ฆ ๐๐โ
โโ (5.10)
Where ๐ is the orthogonal axis of line ๐.
๐ฅ = ๐๐๐๐ ๐โ ๐๐ ๐๐๐ (5.11)
๐ฆ = ๐๐ ๐๐๐+ ๐๐๐๐ ๐ (5.12)
Based on Radon transform [16][17], the estimation of the reference orientation for
the central point ๐ถ is as follows:
Step1) Select the circular neighborhood where the radius = 64 and denotes this
region as a 2-D function ๐ (๐ฅ, ๐ฆ). Subsequently, compute the radon transform
of ๐ (๐ฅ, ๐ฆ) from 0 to2๐ and get ๐ ๐ (p,๐), where ๐ โ 0,2๐ .
Step2) Choose a reference point ๐๐ on the ๐ axis neighborhood ฮฉ โ ๐๐ โ ๐ก, ๐๐ + ๐ก
as shown in figure5.8.
80
The reference orientation ๐C are estimated by
๐C
= arg max๐
๐ ๐ p, ๐ , ๐ โ 0,2๐
๐๐+๐ก
๐=๐๐โ๐ก
(5.13)
Here ๐C is not the extract orientation of the central point ๐ถ. However, it provides us
with a reference orientation, which could be used to calculate the relative difference
in orientations between ๐ถ and other keypoints.
5.6 Experimental Results
In this work, a Euclidean distance metric and Receiver Operating Characteristics
(ROC) are chosen to distinguish the robustness and discriminability of the perceptual
hashing scheme. These are:
a) Euclidean Distance
Let ๐ = ๐ ๐ ๐=1๐ be the set of original images in the database. The corresponding
hashes space H S = H si i=1N where H (๐ ๐) = {๐1 ๐ ๐ , ๐2 ๐ ๐ , โฆ โฆ โฆ๐3(๐ ๐)} is the
hash vector with length n for image ๐ ๐ . Hence, we use Euclidean distance
D( ๐1), (๐2 ) to measure the similarity between two hash vectors H(๐ 1) and H(๐ 2).
Subsequently, given a query image Q, we generate its hash H(Q) and calculate its
distance to each original image in the hash space H(S). The distance between two
hashes is defined as the square root of the sum of the squares of the differences
between the corresponding hash values. i.e., the distance between two hashes
๐1 ๐๐๐ ๐2 is given by
๐๐๐ ๐ก ๐1), (๐2 = ๐1๐โ ๐2๐
2
๐=1
(5.14)
81
b)Receiver Operating Characteristics (ROC)
The ROC curve is used to evaluate the identification performance of the proposed
robust minutiae based fingerprint image hashing technique. To plot the ROC curve,
the ๐๐๐ ๐ and ๐น๐๐ ๐ are defined as:
๐๐๐ ๐ = ๐๐๐๐๐๐๐๐๐๐ก๐ฆ (๐ท ๐ป๐ ๐ผ , ๐ป๐ ๐ผ๐ ) < ๐ (5.15)
๐น๐๐ ๐ = ๐๐๐๐๐๐๐๐๐๐ก๐ฆ (๐ท ๐ป๐ ๐ผ , ๐ป๐ ๐ผโฒ ) < ๐ (5.16)
Where, ๐ is the identification threshold. The image ๐ผ๐ is a modified version of I and
๐ผโฒ is a distinct image of the original image I. ROC curves were generated by varying
the threshold ๐ from the minimum to the maximum value of all the distances. TPR
against FPR were plotted in ROC curves which suggest that the best possible
performance should correspond to point in the top left corner (coordinate 0, 1) of the
ROC space.
Table 5.2: Different Attacks used to assess the hashing performance
Attack
Parameters
Image Processing Operations
JPEG lossy compression
Median filtering
Gaussian Blur
Quality Factor =10
Window size 3x3
Variance ๐ = 0.5, Window size 3x3
Geometric Distortion
Rotation
Translation Degree 50
Variance ๐ = 0.5, Window size 5x5
c) Result and Analysis
The proposed technique was evaluated using 100 images in the FVC 2002/DB1_A
database (Figure 5.9) [115] and we evaluated the perceptual robustness of the image
hashing techniques, RSCH and ASCH against the known attacks, as mentioned in
Table 5.2. The selected length of the hash vector for the RSCH and ASCH is L=20
[14]. The proposed robust feature extraction of fingerprint images, such as SIFT-
Harris-Minutiae, SIFT-Wavelet-Minutiae and SIFT-Minutiae were compared to the
82
SIFT technique, as shown in Figure 5.10- 5.13(a-e). The results are discussed as
follow:
Figure 5.9 Fingerprint images from FVC2002/DB1_A database.
SIFT-Harris-Minutiae: The advantage of generating hashes on robust feature points
lie in the robustness against geometric transform, especially the rotation attacks. The
location to extract robust feature are determined by detected keypoint, the
corresponding hashes are invariant to rotation transform. The proposed fingerprint
image hashing based on SIFT-Harris-Minutiae approaches can achieve better
performance for JPEG lossy compression (QF=10%) and translation ( ๐ = 0.5, 5x5
window) attacks and exhibits good robustness against different attacks especially for
median filtering (3x3 window), Gaussian blur ( ๐ = 0.5, 3x3 window) and a rotation
(5 degrees) as shown in Figure 5.10. It can be seen that the shape contexts provide
an outstanding description of the geometric structure of a shape and allows
embedding the geometric distribution of robust minutiae feature points as well as
their descriptors into a short hash vector. The result also demonstrates that ASCH
relatively outperforms the RSCH. This leads that the distribution of feature points in
the angular direction is better discriminated than in the radial direction [16].
SIFT-Wavelet-Minutiae: Fingerprint image hashing based on an effective
combination of wavelet based feature and SIFT feature along minutiae, to determine
most prominent features for fingerprint image. The experimental result shows that
83
fingerprint image hashing based on SIFT-Wavelet minutiae are provides improved
robustness for the median filtering (3x3 window) and Gaussian blur ( ๐ = 0.5, 3x3
window) and a rotation (5 degrees). However, through RSCH, its hash shown high
sensitive to the JPEG lossy compression (QF=10%) and translation attacks ( ๐ = 0.5,
5x5 window) as shown in Figure 5.11. This is because of RSCH has less capacity in
distribution of feature points in radial direction.
SIFT-Minutiae: The SIFT combined with minutiae based hashing technique
performs moderately well for median filtering (3x3 window), Gaussian blur ( ๐ =
0.5, 3x3 window), although it is highly susceptible to the JPEG lossy compression
(QF=10%), rotation (5 degrees) and translation attacks ( ๐ = 0.5, 5x5 window)
(Figure 5.12).
SIFT: For the sake of comparison the SIFT was compared with the proposed
technique. The performance is slightly affected by the median filtering (3x3
window), and highly sensitive to JPEG lossy compression (QF=10%), Gaussian blur
( ๐ = 0.5, 3x3 window), rotation (5 degrees) and translation attacks ( ๐ = 0.5, 5x5
window) (Figure 5.13). Alongside, To justify the research contribution of combined
methods, experiment was performed on SIFT -Harris and SIFT-wavelet approach.
The performances of these two approaches are highly competitive and resulted in
Table 5.2.
The advantage of generating hashes based on the robust minutiae of fingerprint
image lies in the robustness against geometric distortions like rotations and
translations. Furthermore, the hashing techniques ASCH and RSCH perform better
for identification accuracy. Also, we have observed that Angular shape context
hashing relatively outperforms Radial shape context hashing (Table 5.2), which
indicates that the distribution of feature points in the angular direction is better
discriminative capacity than the distribution in the radial direction. However, image
hashing using feature points still has limitations on considering the distortions of
additive noise and blurring large scale.
84
Table 5.3: Summary of proposed fingerprint image hashing technique
Performance Evaluation of Fingerprint Image Hashing Technique
Attacks
SIFT- Harris -
Minutiae
SIFT- Wavelet-
Minutiae SIFT- Minutiae SIFT SIFT- Harris[16] SIFT- Wavelet
ASCH RSCH ASCH RSCH ASCH RSCH ASCH RSCH ASCH RSCH ASCH RSCH
JPEG lossy
Compression
QF=10%
Good Good Good Low Good Low Moderate Low Good Moderate Moderate Low
Median
filtering 3x3 Better Good Better Moderate Better Moderate Moderate Low Good Good Good Moderate
Gaussian
Blur ฯ=0.5,
3x3
Better Better Better Better Moderate Moderate Moderate Low Better Moderate Good Good
Rotation 50 Better Good Good Moderate Moderate Low Low Low Good Good Moderate Moderate
Translation
ฯ=0.5, 5x5 Good Good Good Low Low Low Low Low Good Moderate Good Low
85
(a) ROC curves under JPEG lossy compression
(b) ROC curves under median filter
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT-Harris-Minutiae
RSCH:SIFT-Harris-Minutiae
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT-Harris-Wavelet
RSCH:SIFT-Harris-Wavelet
86
(c) ROC curves under Gaussian blur
(d) ROC curves under rotation
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT-Harris-Minutiae
RSCH:SIFT-Harris-Minutiae
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT-Harris-Minutiae
RSCH:SIFT-Harris-Minutiae
87
(e) ROC curves under translation
Figure 5.10: ROC curves for the proposed robust minutiae of the fingerprint image
(SIFT-Harris-Minutiae) using the shape context based image hashing technique, (a)
ROC curves under JPEG lossy compression (b) ROC curves under median filter (c)
ROC curves under Gaussian blur (d) ROC curves under rotation (e) ROC curves
under translation
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT-Harris-Minutiae
RSCH:SIFT-Harris-Minutiae
88
(a) ROC curves under JPEG lossy compression
(b) ROC curves under median filter
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH: SIFT-Wavelet-Minutiae
RSCH: SIFT-Wavelet-Minutiae
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH: SIFT-Wavelet-Minutiae
RSCH: SIFT-Wavelet-Minutiae
89
(c) ROC curves under Gaussian blur
(d) ROC curves under rotation
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH: SIFT-Wavelet-Minutaie
RSCH: SIFT-Wavelet-Minutaie
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH: SIFT-Wavelet-Minutaie
RSCH: SIFT-Wavelet-Minutaie
90
(e) ROC curves under translation
Figure 5.11: ROC curves for the proposed robust minutiae of the fingerprint image
(SIFT-Wavelet-Minutiae), using the shape context based image hashing technique,
(a) ROC curves under JPEG lossy compression (b) ROC curves under median filter
(c) ROC curves under Gaussian blur (d) ROC curves under rotation (e) ROC curves
under translation
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH: SIFT-Wavelet-Minutiae
RSCH: SIFT-Wavelet-Minutiae
91
(a) ROC curves under JPEG lossy compression
(b) ROC curves under median filter
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT-Minutiae
RSCH:SIFT-Minutiae
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT-Minutiae
RSCH:SIFT-Minutiae
92
(c) ROC curves under Gaussian blur
(d) ROC curves under rotation
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH: SIFT-Minutiae
RSCH: SIFT-Minutiae
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT-Minutiae
RSCH:SIFT-Minutiae
93
(e) ROC curves under translation
Figure 5.12: ROC curves for the proposed robust minutiae of the fingerprint image
(SIFT-Minutiae) using the shape context based image hashing technique, (a) ROC
curves under JPEG lossy compression (b) ROC curves under median filter (c) ROC
curves under Gaussian blur (d) ROC curves under rotation (e) ROC curves under
translation.
10-3
10-2
10-1
100
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT-Minutiae
RSCH:SIFT-Minutiae
94
(a) ROC curves under JPEG lossy compression
(b) ROC curves under median filter
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH: SIFT
RSCH: SIFT
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT
RSCH:SIFT
95
(c) ROC curves under Gaussian blur
(d) ROC curves under rotation
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH: SIFT
RSCH: SIFT
10-3
10-2
10-1
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT
RSCH:SIFT
96
(e) ROC curves under translation
Figure 5.13: ROC curves for the proposed robust minutiae of the fingerprint
image(SIFT) using the shape context based image hashing technique, (a) ROC
curves under JPEG lossy compression (b) ROC curves under median filter (c) ROC
curves under Gaussian blur (d) ROC curves under rotation (e) ROC curves under
translation.
10-3
10-2
10-1
100
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR
TP
R
ASCH:SIFT
RSCH:SIFT
97
5.7 Summary
In this chapter, we have developed robust minutiae based fingerprint image hashing.
Based on the geometric invariance of SIFT-Harris keypoints, we combined the
minutiae of the fingerprint with the SIFT-Harris feature to detect robust minutiae.
Furthermore, we incorporated the orientation and descriptor in the minutiae of
fingerprint image and fingerprint identification is performed using hashed robust
minutiae. It can be noted that shape contexts provide an outstanding description of
the geometric structure of a shape. Thus, we can embed the geometric distribution of
robust minutiae feature points as well as their descriptors into a short hash vector.
Therefore, SIFT-Harris-Minutiae are more suitable for generating a template and for
comparison of the fingerprint content. In next chapter, the perceptual hashing
technique to improve the minutiae extraction of the fingerprint image is presented.
98
Chapter 6
Perceptual Hashing For Efficient Fingerprint-Based
Identification
In the fingerprint recognition system, the protection of feature points as well as their
authentic usage is an important issue. Problems in image data protection have
emerged due to recent developments in the field of multimedia systems and
communication technologies. Therefore, the security of image data has gained much
more importance and thus, the researcher has been developing approaches for
protecting and authenticating image data.
A prominent solution to protect image data is perceptual hashing. Nevertheless, other
methods, for instance steganography, watermarking and cryptography that can
provide protection for image data exist, alongside perceptual hashing.
Steganography and watermarking both come under data hiding techniques and are
used to hide secret information in the original image. However, differences exists
between steganography and watermarking i.e., steganography conceals the very
existence of secret information.
Therefore, if the existence of secret information is revealed, the steganography fails,
whereas, in watermarking the existence of secret information can be known. The
goal of watermarking is to make the removal/manipulation of secret information
impossible. In contrast, cryptography does not conceal the existence of secret
information; rather it encrypts the information in such a way that it appears useless to
an imposter unless decrypted with an appropriate key. The advantage and properties
of perceptual hashing has been discussed in chapter 5.
In this research, we use perceptual hashing to improve the minutiae extraction of
fingerprint images. We particularly focus on securing the hash and improving
minutiae extraction, even if the fingerprint image has been distorted. The proposed
method extracts the hash after the wavelet transform and singular value
decomposition (SVD). Consequently, the extracted hash obtains good
imperceptibility and robustness. The recent literature describes the robustness and
99
imperceptibility of image data. Lei et al [116] presented robust and reversible
watermarking schemes that embed and /or extract watermarks blindly using
Recursive Dither Modulation (RDM) with a combination of wavelet transform and
Singular Value Decomposition (SVD). In addition, Differential Evolution (DE)
optimization is used to control the strength of the watermark. This method
demonstrates excellent robustness and imperceptibility, in terms of Structural
Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR).
Additionally, Aslantas [117] presented a new optimal method of robust image
watermarking based on SVD using DE. The DE was employed to optimise the
fitness function to achieve maximum robustness and transparency.
Bhatnagar and Raman [118] presented a new semi-blind reference watermarking
scheme based on Discrete Wavelet Transform (DWT) and Singular Value
Decomposition (SVD) for copyright protection and authenticity. The result
demonstrates the semi-blind reference watermarking scheme withstand to various
noise operation like average fingering, median filtering, additive Gaussian noise,
JPEG compression, cropping, rotation and other blurring operation.
Ghouti et al [119] presented a novel image-adaptive watermarking scheme based on
a subband decomposition to balance multiwavelet transform and this scheme use
various statistical models for the host image to derive the watermarking (data-hiding)
capacity.
Ramakrishnan et al [120] developed a hybrid image watermarking algorithm which
satisfies both imperceptibility and robustness requirements. The watermarking
scheme use singular values of Wavelet Transformationโs HL and LH subbands to
embed watermark. The efficiency of the scheme are explained through PSNR,
Normalized Cross Correlation (NCC) and gain factor against to signal and image
processing operation.
Moreover, Hore and Ziou [121] analysed a theoretical study to compare the PSNR
and SSIM. The study reveals that an analytical link exists between them, which
work for various kinds of image degradations, such as Gaussian blur, additive white
noise, jpeg and jpeg 2000 compression. The authors also explained the sensitivity of
PSNR and SSIM to these degradations
100
Tang et al [122] demonstrated a metric to evaluate the perceptual similarity between
an original image and its distorted version. The result discloses that the metric is
insensitive to content-preserving processing, for instance JPEG compression,
rotation, low-pass filtering, watermarking embedding and other moderate noise, but
that it is very sensitive to malicious modification. The metric evaluation is useful in
applications, such as image hashing and content-based image retrieval.
Image quality assessment plays an important role in the field of image manipulation.
Wang et al [123] summarized the traditional approach to image quality assessment
and its limitations. The authors introduced structural similarity as an alternative
principle for the design of image quality measures. Al-Najjar et al [124] presented
the comparison of image quality assessment between PSNR, HVS, SSIM and
universal image quality index (UIQI) metrics.
Additionally, Hofbauer et al [125] described fusion scenario by combining image
metrics and hamming distance approaches in iris biometric systems. The
incorporation of distinct image metrics in a fusion scenario significantly improves
the recognition accuracy of systems. Keimel and Diepold [126] improved the
predication accuracy of PSNR by simple temporal pooling and demonstrate the
effectiveness of temporal pooling on a set of high โdefinition television sequences
and broadcasting applications, whilst Run et al [127] demonstrated two methods to
develop reliability and robustness: the principle components of the watermark are
embedded into the host image in discrete cosine transform (DCT) and inserted into
the host image in DWT. The particle swarm optimization (PSO) is used for suitable
scaling factors, to improve robustness.
In addition, Braeckman et al [128] illustrated a flexible framework algorithm for
reduced reference (RR) visual quality assessment, which is based on perceptual
hashing and image and video watermarking. The RR quality assessment is based on
perceptual hashing and watermarking techniques. Gao et al [129] proposed a
reduced-reference image quality assessment framework, by incorporating merits of
multiscale geometry analysis, contrast sensitivity function and Weberโs law of just-
noticeable difference.
101
Li et al [130] recommended a method of effective combination of the human vision
system with regional mutual information (RMI). This method measures the similarity
between an original image and its distorted version. The performance comparison
shows that the MRMI method performs better than the SSIM and PSNR method.
Weng et al [131] suggested a novel image authentication system by combing
perceptual hashing and robust watermarking. In these algorithms, an image is divided
into blocks and each block is represented by a compact hash value, which is
embedded in the block. The authenticity of the image can be verified by re-
computing hash values and comparing them with the ones extracted from the image.
Thus, the system tolerates a wide range of attacks and tamper location of the image.
6.1 Singular value decomposition (SVD)
The basic concepts of singular value decomposition (SVD) operations are discussed
as follows: SVD decomposes a ๐x๐ real matrix A into a product of 3 matrices:
๐ด = ๐๐๐๐ = ๐ข1, ๐ข2, ๐ข3โฆโฆ.,๐ข๐
๐1
๐2
๐3
๐ฃ1, ๐ฃ2, ๐ฃ3โฆโฆ.,๐ฃ๐ ๐ (6.1)
Where s is a ๐x๐ diagonal matrix ๐ and ๐๐are ๐x๐ orthogonal matrices, whose
column vectors are ๐ข๐ โฒ๐ and ๐ฃ๐ โฒ๐ , respectively. The elements of ๐ are only non-Zero
on the diagonal arranged in decreasing order and are called the ๐๐๐ of A. when the
rank of A is r, S = diag(ฯ1, ฯ2, โฆ โฆ , ฯn) satisfiesฯ1 โฅ ฯ2 โฅ โฏ โฅ ฯr โฅ ฯr+1 =
ฯr+2 โฆ โฆ ฯn = 0. Let A be a matrix whose elements are pixel values of an image.
The image can be written as:
๐ด = ฯi
๐
๐=0
๐ข๐๐ฃ๐๐
(6.2)
102
6.1.1 Properties of SVD
SVD has several attractive properties from the viewpoint of image processing
applications. SVD efficiently represents the intrinsic algebraic properties of an
image, where singular values specify the brightness of the image and the
corresponding pair of singular vectors reflects the geometry of the image. The main
properties of SVD operations are as follows:
Stability: Let๐ด,๐ต โ ๐ ๐๐ฅ๐ and their corresponding SVs are ฯ1, ฯ2, โฆ โฆ , ฯn
and ฯ1 , ฯ2 , โฆ โฆ , ฯn , respectively. Then a relation can be established between them
as ฯi โ ฯi โค ๐ด โ ๐ต 2. This indicates that the singular value of an image has
very good stability.
Proportionality: The singular values of ๐ด(ฯ1, ฯ2, โฆ โฆ , ฯn) and the singular values
of ๐๐ด(ฯ1,โ ฯ2
โ , โฆ โฆ โฆ . ฯ๐โ ) are related as ๐ (ฯ1, ฯ2, โฆ โฆ , ฯn)=(ฯ1,
โ ฯ2โ , โฆ โฆ โฆ . ฯ๐
โ )
which indicates that the proportion invariance of singular value must depend on
standardization of singular value.
Transpose: ๐ด and its transposed counterpart ๐ด๐ have the same non-zero singular
values.
Flip: ๐ด and its flipped versions give the same non-zero singular values.
Rotation: ๐ด and its rotated version obtained by rotating ๐ด through an arbitrary
angle have the same non-zero singular values.
Scaling: Let ๐ด โ ๐ ๐๐ฅ๐ has the singular values ฯ1, ฯ2, โฆ โฆ , ฯn then its scaled
counterpart is equal to ๐ด๐ and has singular values equal to ฯ๐โ ๐ฟ๐๐ฟ๐ , where ๐ฟ๐ and
๐ฟ๐ are the scaling factors or rows and columns, receptively. If the scaling function
affects only rows or columns, ๐ด๐ has the singular values equal to ฯ๐โ ๐ฟ๐ or ๐ฟ๐ ,
respectively.
Translation: Both the matrix ๐ด and its translated counterpart ๐ด๐ก have the same
non-zero singular values.
6.2. Proposed hash security for fingerprint images
Figure 6.1 illustrates the block diagram for the proposed hash security technique of
the fingerprint image, which is presented in detail below:
103
Step1: The original fingerprint image (๐ผโ) is partitioned into sub-blocks ๐ตโ๐ ,
where ๐ = 1,2, โฆ . ๐.
Step 2: Two-level transform is performed to each subblock ๐ตโโ๐ . The approximate
coefficients are selected for SVD calculation. The basic operation of SVD
and its properties are explained in section 6.1 and 6.1.1 respectively.
Step 3: Perform SVD on the low frequency of each block to generate SVs ๐โโ๐ .
Step 4: The SVs (๐โโ๐ ) is normalised as
๐๐โ๐พ= ๐โโ๐ (6.3)
Let ๐๐โโ๐พ=floor ๐๐
โโ๐พ /โ๐พ and the hash is extracted with the following rule
๐ธ โ ๐, ๐ = 0, ๐๐๐ ๐๐ค
๐พโ, 2 = 0
1, ๐๐๐ ๐๐ค๐พโ, 2 = 1
(6.4)
Where floor (. ) is rounding toward the negative infinity and โ๐พ is quantization steps
Step 5: The final hash is obtained from the original fingerprint image.
Original Image
Has
h
Noise AdditionAttacked Image
Image Correction
Enhanced/ Proposed Image
Hash Extraction
Figure 6.1: Proposed hash securing for fingerprint images
The hash is extracted after wavelet transform and singular value decomposition and
stored in the database. As shown in the figure 6.1, when the given image is distorted
by noise (Table6.1), the stored hash is used to correct the distorted image. Thus, the
corrected image (Enhanced/Proposed Image) is close to the original image in
structural similarity. Through this approach, we can retain the maximum minutiae of
the fingerprint image, even if the original image is distorted as mentioned in Table
6.1. The performance evaluation of the proposed approach includes important
104
metrics, such as SSIM and PSNR, which are utilized to represent imperceptibility.
Experimentally, the proposed technique demonstrates good quality robustness
against image processing operations and geometric attacks. The results are detailed
in section 6.4.
6.3 Performance Evaluation of Proposed Approach
The proposed approached is evaluated in relation to their robustness and
imperceptibility. The SSIM and PSNR are utilized to represent imperceptibility, and
BER used to measure robustness. These are explained in the following section.
6.3.1 Structural similarity Index Measure (SSIM)
The structural similarity index measure (SSIM) is used to measure quality by
capturing the similarity of images [123] based on three comparisons: luminance,
contrast and structure, which are selected for the measure of imperceptibility. Three
components are combined to yield an overall similarity measure as:
SSIM ๐, ๐ = ๐ ๐, ๐ ๐ ๐, ๐ ๐ ๐, ๐ (6.5)
๐ ๐, ๐ =2๐๐ ๐๐ +๐ถ1
๐๐2 +๐๐
2 +๐ถ1
๐ ๐, ๐ =2๐๐ ๐๐ +๐ถ2
๐๐2 +๐๐
2+๐ถ2
๐ ๐, ๐ =๐๐๐ +๐ถ3
๐๐ ๐๐ +๐ถ3
(6.6)
The first term in the equation (6.6) is the luminance comparison function, which
measures the closeness of the two images mean luminance ๐๐and๐๐ . The second
term ๐ ๐, ๐ is the contrast comparison function, which measures the closeness of
the contrast of the two images. Consequently, the contrast is measured by the
standard deviation ๐๐ and ๐๐ . The third term๐ ๐, ๐ is the structure comparison
function, which measures the correlation coefficient between the two
images ๐and ๐. The covariance between ๐ and ๐ is represented as ๐๐๐ . The positive
value of the SSIM index is in [0, 1]. A value of 0 means no correlation between the
images, and 1 means that ๐ = ๐. The positive constant ๐ถ1 , ๐ถ2 , ๐ถ3 provides stability.
105
By combining the three comparison functions, the SSIM index is obtained as
SSIM ๐, ๐ = ๐ ๐, ๐ ๐ผ . ๐ ๐, ๐ ๐ฝ . ๐ ๐, ๐ ๐พ (6.7)
Where ๐ผ> 0, ๐ฝ> 0 and ๐พ > 0 are parameters used to adjust the relative importance of
the three components and the parameters are set as ๐ผ = ๐ฝ = ๐พ and ๐ถ3=๐ถ2/2. From
the above parameters, the SSIM index can be defined as:
SSIM ๐, ๐ = 2๐๐ ๐๐ +๐ถ1 (2๐๐๐ +๐ถ2)
(๐๐2 +๐๐
2 +๐ถ1)(๐๐2 +๐๐
2+๐ถ2) (6.8)
6.3.2 Peak Signal-to-Noise Ratio (PSNR)
Given a reference image X and a test image Y, both sizes MxN, the PSNR between,
the PSNR between X and Y is defined as:
PSNR ๐, ๐ = 10log10 2552
๐๐๐ธ ๐, ๐ (6.9)
Where,
๐๐๐ธ ๐, ๐ = 1
๐๐ ๐๐๐ โ ๐๐๐
2๐
๐=1
๐
๐=1
(6.10)
The PSNR [114] value approaches infinity as the MSE approaches zero; therefore,
this shows that a higher PSNR value provides an enhanced quality image. At the
other end of the scale, a small value of the PSNR implies a high numerical difference
between images.
6.3.3 Bit- Error- Rate (BER)
In digital transmission, the number of bit errors is the number of received bits of a
data stream over a communication channel that have been altered due to noise,
interference, distortion or bit synchronization errors. The bit error rate or bit error
ratio (BER) is the number of bit errors divided by the total number of transferred bits
106
during a studied time interval. The BER is a unitless performance measure, often
expressed as a percentage number.
BER = Errors /Total Number of Bits (6.11)
6.4 Experimental Results
The image hashing scheme are applied in the fields of image authentication and
retrieval. In general the experiments are tested on 20 standard natural images to show
the robustness against usual content preserving operations [132]. The proposed hash
security technique is applied on 20 randomly selected images from fingerprint
database of FVC2002/DB1_A. The number of blocks has a great effect on the
extraction of hash and is used to correct the image. Therefore, the number of blocks
should be taken in such a way so as to achieve a superior PSNR and SSIM. In this
research, the block size is 8x8. In addition, the performance evaluation of the
proposed approach includes important metrics, such as robustness and
imperceptibility. Hence, the SSIM and PSNR are utilized to represent
imperceptibility, while the BER is used to measure robustness.
The PNSR is easy to evaluate but is not always in accordance with human judgment
of quality. On other hand, the SSIM is closer to human visual system. It can be seen
that the proposed hash security technique is tested on content-preserving operations
(Table 6.1): rotation (1 degree), average filter (3x3 window), Gaussian noise (๐ =
0.005), salt and pepper (๐ = 0.01), JPEG compression of 90%, and Contrast Low
(CL) and Contrast High (CH). Table 6.2 gives the average SSIM and PSNR values
between the proposed image and distorted image. The wavelet transform- singular
value decomposition (DWT-SVD) method is included for comparison.
A drawback of basic SSIM index [133] has sensitivity to relative operation like
rotation, JPEG compression, Gaussian noise, salt and pepper noise, blurring,
translation and scaling. The proposed hash security technique provides high SSIM
values (closer to value of 1) for average filter (3x3 window), JPEG compression
(QF=90%), Contrast Low (CL) and slightly sensitive to rotation (1 degree), salt and
pepper (๐ = 0.01) and Contrast High (CH), whilst Gaussian noise (๐ = 0.005) has
better correlation when compared to DWT-SVD method.
107
The PSNR improvement for the proposed image are more than 3dB for rotation (1
degree), average filter (3x3 window), Gaussian noise (๐ = 0.005), JPEG
compression (QF=90%) and Contrast Low (CL). The PSNR value is less sensitivity
for salt and pepper (๐ = 0.01) and Contrast High (CH), ranging from 2.5 dB to 3dB.
The average SSIM and PNSR values for different attacks are illustrated in figures 6.2
and 6.3. The plot demonstrates the metric performance to evaluate the perceptual
similarity between a proposed image and its distortion version. The result reveals that
the metric achieves a higher value of SSIM and PSNR values for the fingerprint
images for all the content-preserving operations.
To assess the robustness, the BER was calculated between the original and enhanced
image. The result shows that the proposed method demonstrates good robustness.
However, the approached method is still sensitive to a rotation (1 degree) and
Gaussian noise (๐ = 0.005). The robustness moderately affected the average filter
(3x3 window), JPEG lossy compression (QF= 90%), and the Contrast Low (CL) and
Contrast High (CH). Hence, for salt and pepper noise (๐ = 0.01), the bit error rate is
good.
In addition, we observed the performance of the minutiae extraction technique as
shown in Table 6.3 and Figure 5.4. Subsequently, we calculate the Hausdorff
distance (as explained in section 4.1.1) between the minutiae from the distorted
image and the enhanced image. The Hausdorff distance value is comparatively low
for our proposed approach, by retaining maximum minutiae compared to the DWT
SVD method. Table 6.4 describes the performance evaluation of the metrics and
minutiae of the fingerprint images.
108
Table 6.1: Content preserving operation used to assess the hash performance
Attack Parameters
Image Processing Operations
JPEG lossy compression
Gaussian Noise
Average Filter
Salt and Pepper
Contrast Low
Contrast High
Quality Factor =90%
Variance = 0.005
Window size 3x3
Variance = 0.01
/
/
Geometric Distortion
Rotation
Degree 10
109
Table 6.2: Average SSIM and PSNR Value
Attack and its
Parameters
Average SSIM Value
Average PSNR Value
Average
BER
Attacked
Image
Proposed
Image
Difference
between the
attacked and
proposed
image
Attacked
Image
Proposed
Image
Difference
between the
attacked and
proposed
image
Rotation 10 0.6825 0.8306 0.1481 16.6131 19.9298 3.3167 0.3209
Average Filter 3x3 0.8615 0.9345 0.073 22.8825 25.8975 3.015 0.3089
Gaussian Noise
V=0.005 0.4968 0.5543 0.0575 24.5433 27.6197 3.0764 0.5074
Salt and pepper
V=0.01 0.7129 0.8026 0.0897 23.7273 26.2882 2.5609 0.1871
Jpeg 90% 0.9164 0.9531 0.0367 40.6526 43.6505 2.9979 0.3002
Contrast Low(CL) 0.9179 0.9499 0.032 19.3134 22.3275 3.0141 0.2924
Contrast High(CH) 0.8540 0.8638 0.0098 19.2542 22.2454 2.9912 0.2982
110
Table 6.3: Average Hausdorff Distance of the Minutiae
Attack and its Parameters
Average Hausdorff
distance
Attacked
Image
Proposed
Image
Rotation 10 0.0103 0.0106
Average Filter 3x3 0.0161 0.0112
Gaussian Noise V=0.005 0.0133 0.0107
Salt and pepper V=0.01 0.0201 0.0062
Jpeg 90% 0.0053 0.0024
Contrast Low(CL) 0.0548 0.0117
Contrast High(CH) 0.0282 0.0131
Table 6.4: Performance Evaluation of Metrics and Minutiae
Attacks
SSIM
PSNR
Hausdorff distance
of the Minutiae
Rotation 10 Better Good Slightly sensitivity
Average Filter 3x3 Good Good Better
Gaussian Noise
V=0.005 Moderate Good Better
Salt and pepper
V=0.01 Better Good Good
Jpeg 90% Good Good Good
Contrast Low(CL) Good Good Good
Contrast High(CH) Better Good Good
111
Figure 6.2: Average SSIM
Figure 6.3: Average PSNR
Figure 6.4: Average Hausdorff Distance of Minutiae
112
6.5 Summary
In this research, a robust perceptual hash extract technique using DWT-SVD method
is proposed. The proposed method focus on extraction and securing the hash in the
database after wavelet-SVD combination. The tradeoff between robustness and
imperceptibility is achieved by proper selection of quantization steps. Overall, the
proposed method demonstrates to enhance the fingerprint image and provide good
balance of robustness and imperceptibility. Additionally, this approach retains the
maximum minutiae of the fingerprint image, even if the given image is distorted.
113
Chapter 7
Conclusion and future work
7.1 Introduction
With the rapid increase in biometric recognition systems in the commercial sector,
the security of stored biometric data is increasingly becoming crucial. Current
biometric systems have a number of vulnerabilities and a motivated adversary can
undoubtedly cause severe harm to a biometric system, as well as to users enrolled in
the system. Furthermore, due to the permanent nature of biometrics data, its theft and
misuse may be irreversible and have lasting consequences.
This thesis focuses on improving the security of fingerprint templates and accurate
comparison of fingerprint content. The generation of fingerprint templates, which in
turn are used to compare fingerprint content or their perceptual hashes mostly rely on
feature extraction techniques, for instance SIFT or fingerprint minutiae. We used
shape context based perceptual hashes using the RSCH and ASCH methods to plot
the accuracy of fingerprint comparison using ROC curves. The framework of
perceptual fingerprint image hashing algorithms has major components, including
pre-processing on image, feature extraction and post processing. The robustness of
fingerprint image hashing in conjunction with content-preserving operations such as
Gaussian noise, JPEG compression, filtering and geometric attacks are investigated.
7.2 Contribution of the thesis
The design of fingerprint image hashing has three important modules: feature
extraction, feature compression and security issues. In this thesis, I focused mainly
on robust feature extraction and security issues. The contributions of this thesis are
concluded as follows:
We discuss the implementation of robust minutiae based fingerprint image
hashing. In this method, the minutiae of the fingerprint is combined with the
SIFT-Harris features (that tolerate geometric invariance) to detect robust minutiae.
114
In addition, the orientation and descriptor in the minutiae of the fingerprint image
is incorporated whilst fingerprint identification is performed using hashed robust
minutiae.
It shows that the shape contexts provide an outstanding description of the
geometric structure of a shape. Moreover, this approach allows embedding the
geometric distribution of robust minutiae feature points as well as their descriptors
into a short hash vector. The experimental results demonstrate that angular shape
context hashing relatively outperforms the radial shape context hashing. This
leads to conclusion that the distribution of feature points in the angular direction is
better discriminated than in the radial direction, however with slightly sensitivity
to the geometric attacks.
The important contribution is fingerprint image hashing based on an effective
combination of wavelet based feature and SIFT feature. In this approach, SIFT-
Wavelet is combined with the Fingerprint Minutiae extraction method to
determine the most prominent fingerprint features. These features are post-
processed using RSCH and ASCH techniques to plot the accuracy of fingerprint
comparison using ROC curves.
It shows that SIFT-Wavelet minutiae based fingerprint image hashing is robust to
attacks, such as median filter and Gaussian blur and rotation, but high sensitivity
to the JPEG compression and translation. For further investigation, SIFT is
combined with the Fingerprint Minutiae extraction procedure and post-processed
using RSCH and ASCH techniques to plot the accuracy. The experimental results
demonstrate that the fingerprint template and accuracy of fingerprint comparison
improved when a combination of two different feature extraction techniques are
used, in contrast to using only a single feature extraction procedure.
ROC plots of SIFT-Harris-Minutiae, SIFT-Wavelet-Minutiae, SIFT-Minutiae and
SIFT are shown in Chapter 4. These ROCs further demonstrate that the SIFT-
Harris-Minutiae outperform other combinations. Therefore, the new SIFT-Harris-
Minutiae technique is more suitable for generating a template and comparison of
the fingerprint content.
115
The third contribution of thesis is to enhance the performance of a fingerprint
system, perceptual hashing is used to improve the minutiae extraction of
fingerprint images. The proposed work focused on securing the hash and
improving minutiae extraction, under various content preservation operations. The
hash extraction is performed after wavelet transform and singular value
decomposition (SVD). The performance evaluation of this approach includes
important metrics such as SSIM and PSNR. Experimentally, it has shown
robustness against image processing operations and geometric attacks.
7.3 Future work
This section highlights the direction of future research to improve security of
biometric recognition systems based on hash techniques.
In this research, we developed robust minutiae based fingerprint image
hashing and investigated their perceptual robustness against content
persevering manipulations. Based on the geometric invariance of SIFT-Harris
keypoints, we combined the minutiae of fingerprint with SIFT-Harris feature
to detect robust minutiae. Shape contexts provide an outstanding description
of the geometric structure of a shape. We can embed the geometric
distribution of robust minutiae feature points as well as their descriptors into
a short hash vector and proposed minutiae based fingerprint image hashing,
i.e., The RSCH in the radial direction and ASCH in the angular direction. To
achieve better perceptual robustness, a joint RSCH-ASCH can be promising
future direction for fingerprint image hashing.
The shape contexts based hashes has an advantage of embedding the
geometric structure of the image, the proposed minutiae based fingerprint
image hashing technique could be applied in image tampering detection.
In this research, the robust points in the SIFT have been determined through
approximation coefficients using the Haar wavelet. Horizontal, vertical and
diagonal coefficients are utilized to localize the salient points. A detailed
116
study of various wavelet transform could be conducted to optimise the
performance of fingerprint image hashing.
We presented a technique to improve the minutiae extraction of fingerprint
images using perceptual hashing. The hash is extracted using a combination
of the wavelet algorithm and SVD technique. It demonstrated that the
proposed scheme has superior robustness against image processing operations
and geometric attacks. Machine learning technique will be used to model
these attacks in order to improve the hashing system.
Further the proposed work can be extended to other biometric traits (e.g. face,
palmprint, iris, etc.,) and other image media to enhance recognition
performance. Although biometrics may be susceptible to false matches,
possibly due to scanning and sensor errors, there are ways to minimize this,
currently, by utilizing multi-factor options like a password or smartcard
combined with biometrics to add an extra layer of security towards
authentication. If used together, and not alternatively, the systems are
significantly stronger than when used individually.
Recent trends, multi-biometrics (the use of different sets of biometric data
simultaneously) as a good alternative to increase matching accuracy for
identification and verification. Multimodal biometrics systems, which use
multiple sensors for data acquisition, offer multiple recognition algorithms
and take advantage of each biometric technology while overcoming the
limitations of a single technology.
In terms of use, the future of biometrics could be in mobile devices and
applications for eGovernment, eHealth and eBanking. Through biometric
mobile scanning devices, authentication and identification can be brought to
the field. It is easy to imagine the possible uses for such systems for other
professions, like law enforcement, borders control, medical and emergency
services, or even to secure access to government or financial services.
117
Biometric technology could soon become mainstream to the growth of the
mobile devices market. Biometrics Research Group, Inc.estimates that the
sale of smartphones, in the U.S. only, will grow to 121 million in 2018.
Due to this proliferation and to the increased functionalities they offer
their users, their analysts believe there will be a strong push toward the
integration of biometric technology to replace traditional authentication via
pin and password. Biometrics Research Group, Inc. predicted that already in
2014 over 90 million smartphones would be shipped with biometric
technology, while Goode Intelligence has forecasted that by 2019 the number
of mobile and wearable biometric technology users in the world will reach
5.5 billion.
118
Bibliography
[1] Bouridane, A. (2009). Imaging for Forensics and Security: From Theory to
Practice (Vol. 106). Springer.
[2] Jain, A. K., Ross, A., & Prabhakar, S. (2004). An introduction to biometric
recognition. Circuits and Systems for Video Technology, IEEE Transactions
on, 14(1), 4-20.
[3] Jain, A. K., & Kumar, A. (2010). Biometrics of next generation: An
overview. Second Generation Biometrics.
[4] Ross, A. A. (2003). Information fusion in fingerprint authentication (Doctoral
dissertation, Michigan State University).
[5] Maltoni, D., Maio, D., Jain, A. K., & Prabhakar, S. (2009). Handbook of
fingerprint recognition. Springer.
[6] Jain, A. K., Flynn, P., & Ross, A. A. (2007). Handbook of biometrics.
Springer Science & Business Media.
[7] Jain, A. K., Nandakumar, K., & Nagar, A. (2008). Biometric template
security. EURASIP Journal on Advances in Signal Processing, 2008, 113.
[8] Jain, A. K., Ross, A., & Uludag, U. (2005, September). Biometric template
security: Challenges and solutions. In Proceedings of European Signal
Processing Conference (EUSIPCO) (pp. 469-472).
[9] Ratha, N. K., Connell, J. H., & Bolle, R. M. (2001, January). An analysis of
minutiae matching strength. In Audio-and Video-Based Biometric Person
Authentication (pp. 223-228). Springer Berlin Heidelberg.
[10] Jain, A. K., Feng, J., & Nandakumar, K. (2010). Fingerprint Matching.
Computer, 43(2), 36-44.
[11] Nagar, A., Nandakumar, K., & Jain, A. K. (2010). A hybrid biometric
cryptosystem for securing fingerprint minutiae templates. Pattern
Recognition Letters, 31(8), 733-741.
[12] Han, S. H., & Chu, C. H. (2010). Content-based image authentication: current
status, issues, and challenges. International Journal of Information Security,
9(1), 19-32.
[13] Ahmed, F., Siyal, M. Y., & Uddin Abbas, V. (2010). A secure and robust
hash-based scheme for image authentication. Signal Processing, 90(5), 1456-
1470.
119
[14] Monga, V., & Evans, B. L. (2006). Perceptual image hashing via feature
points: performance evaluation and tradeoffs. Image Processing, IEEE
Transactions on, 15(11), 3452-3465.
[15] Bay, H., Tuytelaars, T., & Van Gool, L. (2006). Surf: Speeded up robust
features. In Computer VisionโECCV 2006 (pp. 404-417). Springer Berlin
Heidelberg.
[16] Lv, X., & Wang, Z. J. (2012). Perceptual image hashing based on shape
contexts and local feature points. Information Forensics and Security, IEEE
Transactions on, 7(3), 1081-1093.
[17] Lefebvre, F., Macq, B., & Legat, J. D. (2002, September). RASH: Radon soft
hash algorithm. In Signal Processing Conference, 2002 11th European (pp.
1-4). IEEE.
[18] Tuyls, P., ล koriรง, B., & Kevenaar, T. (2007). Security with noisy data: on
private biometrics, secure key storage and anti-counterfeiting. Springer.
[19] Moujahdi, C., Bebis, G., Ghouzali, S., & Rziza, M. (2014). Fingerprint shell:
Secure representation of fingerprint template. Pattern Recognition Letters,
45, 189-196.
[20] Liang, X., Bishnu, A., & Asano, T. (2007). A robust fingerprint indexing
scheme using minutia neighborhood structure and low-order delaunay
triangles. Information Forensics and Security, IEEE Transactions on, 2(4),
721-733.
[21] Jin, Z., Jin Teoh, A. B., Ong, T. S., & Tee, C. (2012). Fingerprint template
protection with minutiae-based bit-string for security and privacy preserving.
Expert Systems with Applications, 39(6), 6157-6167.
[22] Liu, E., Zhao, H., Liang, J., Pang, L., Chen, H., & Tian, J. (2012). Random
local region descriptor (RLRD): A new method for fixed-length feature
representation of fingerprint image and its application to template protection.
Future Generation Computer Systems, 28(1), 236-243.
[23] Feng, J. (2008). Combining minutiae descriptors for fingerprint matching.
Pattern Recognition, 41(1), 342-352.
[24] Cappelli, R., Ferrara, M., & Maltoni, D. (2010). Minutia cylinder-code: A
new representation and matching technique for fingerprint recognition.
Pattern Analysis and Machine Intelligence, IEEE Transactions on, 32(12),
2128-2141.
120
[25] Tulyakov, S., Farooq, F., & Govindaraju, V. (2005). Symmetric hash
functions for fingerprint minutiae. In Pattern Recognition and Image Analysis
(pp. 30-38). Springer Berlin Heidelberg.
[26] Tulyakov, S., Farooq, F., Mansukhani, P., & Govindaraju, V. (2007).
Symmetric hash functions for secure fingerprint biometric systems. Pattern
Recognition Letters, 28(16), 2427-2436.
[27] Sutcu, Y., Sencar, H. T., & Memon, N. (2007, April). A geometric
transformation to protect minutiae-based fingerprint templates. In Defense
and Security Symposium (pp. 65390E-65390E). International Society for
Optics and Photonics.
[28] Shuai, X., Zhang, C., & Hao, P. (2008, December). Fingerprint indexing
based on composite set of reduced SIFT features. In Pattern Recognition,
2008. ICPR 2008. 19th International Conference on (pp. 1-4). IEEE.
[29] Xu, H., Veldhuis, R., Bazen, A. M., Kevenaar, T. A., Akkermans, T. A., &
Gokberk, B. (2009). Fingerprint verification using spectral minutiae
representations. Information Forensics and Security, IEEE Transactions on,
4(3), 397-409.
[30] Xu, H., Veldhuis, R., Kevenaar, T. A., & Akkermans, T. A. (2009). A fast
minutiae-based fingerprint recognition system. Systems Journal, IEEE, 3(4),
418-427.
[31] Feng, J., Ouyang, Z., & Cai, A. (2006). Fingerprint matching using ridges.
Pattern Recognition, 39(11), 2131-2140.
[32] Jain, A. K., Chen, Y., & Demirkus, M. (2007). Pores and ridges: high-
resolution fingerprint matching using level 3 features. Pattern Analysis and
Machine Intelligence, IEEE Transactions on, 29(1), 15-27.
[33] Ahn, D., Kong, S. G., Chung, Y. S., & Moon, K. Y. (2008, May). Matching
with secure fingerprint templates using non-invertible transform. In Image
and Signal Processing, 2008. CISP'08. Congress on (Vol. 2, pp. 29-33).
IEEE.
[34] Lee, C., Choi, J. Y., Toh, K. A., & Lee, S. (2007). Alignment-free cancellable
fingerprint templates based on local minutiae information. Systems, Man, and
Cybernetics, Part B: Cybernetics, IEEE Transactions on, 37(4), 980-992.
[35] Chang, S. H., Cheng, F. H., Hsu, W. H., & Wu, G. Z. (1997). Fast algorithm
for point pattern matching: invariant to translations, rotations and scale
changes. Pattern recognition, 30(2), 311-320.
121
[36] Jiang, X., & Yau, W. Y. (2000). Fingerprint minutiae matching based on the
local and global structures. In Pattern Recognition, 2000. Proceedings. 15th
International Conference on (Vol. 2, pp. 1038-1041). IEEE.
[37] Luo, X., Tian, J., & Wu, Y. (2000). A minutiae matching algorithm in
fingerprint verification. In Pattern Recognition, 2000. Proceedings. 15th
International Conference on (Vol. 4, pp. 833-836). IEEE.
[38] Bhowmick, P., Bishnu, A., Bhattacharya, B. B., Kundu, M. K., Murthy, C.
A., & Acharya, T. (2005). Determination of minutiae scores for fingerprint
image applications. International journal of image and graphics, 5(03), 537-
571.
[39] Jain, A., Hong, L., & Bolle, R. (1997). On-line fingerprint verification.
Pattern Analysis and Machine Intelligence, IEEE Transactions on, 19(4),
302-314.
[40] Monga, V., & Mฤฑhรงak, M. K. (2007). Robust and secure image hashing via
non-negative matrix factorizations. Information Forensics and Security, IEEE
Transactions on, 2(3), 376-390.
[41] Wu, M., Mao, Y., & Swaminathan, A. (2007, August). A signal processing
and randomization perspective of robust and secure image hashing. In
Statistical Signal Processing, 2007. SSP'07. IEEE/SP 14th Workshop on (pp.
166-170). IEEE.
[42] Kim, C. (2003). Content-based image copy detection. Signal Processing:
Image Communication, 18(3), 169-184.
[43] Venkatesan, R., Koon, S. M., Jakubowski, M. H., & Moulin, P. (2000).
Robust image hashing. In Image Processing, 2000. Proceedings. 2000
International Conference on (Vol. 3, pp. 664-666). IEEE.
[44] Lefรจbvre, F., Czyz, J., & Macq, B. (2003, September). A robust soft hash
algorithm for digital image signature. In Image Processing, 2003. ICIP 2003.
Proceedings. 2003 International Conference on (Vol. 2, pp. II-495). IEEE.
[45] Swaminathan, A., Mao, Y., & Wu, M. (2006). Robust and secure image
hashing. Information Forensics and Security, IEEE Transactions on, 1(2),
215-230.
[46] Khelifi, F., & Jiang, J. (2010). Perceptual image hashing based on virtual
watermark detection. Image Processing, IEEE Transactions on, 19(4), 981-
994.
122
[47] Khelifi, F., & Jiang, J. (2010). Analysis of the security of perceptual image
hashing based on non-negative matrix factorization. Signal Processing
Letters, IEEE, 17(1), 43-46.
[48] Kozat, S. S., Venkatesan, R., & Mihรงak, M. K. (2004, October). Robust
perceptual image hashing via matrix invariants. In Image Processing, 2004.
ICIP'04. 2004 International Conference on (Vol. 5, pp. 3443-3446). IEEE.
[49] Badrinath, G. S., Gupta, P., & Mehrotra, H. (2013). Score level fusion of
voting strategy of geometric hashing and SURF for an efficient palmprint-
based identification. Journal of real-time image processing, 8(3), 265-284
[50] Tuytelaars, T., & Mikolajczyk, K. (2008). Local invariant feature detectors: a
survey. Foundations and Trendsยฎ in Computer Graphics and Vision, 3(3),
177-280.
[51] Lowe, D. G. (2004). Distinctive image features from scale-invariant
keypoints. International journal of computer vision, 60(2), 91-110.
[52] Koval, O., Voloshynovskiy, S., Bas, P., & Cayre, F. (2009, February). On
security threats for robust perceptual hashing. In IS&T/SPIE Electronic
Imaging (pp. 72540H-72540H). International Society for Optics and
Photonics.
[53] Tuyls, P., Akkermans, A. H., Kevenaar, T. A., Schrijen, G. J., Bazen, A. M.,
& Veldhuis, R. N. (2005, January). Practical biometric authentication with
template protection. In Audio-and Video-Based Biometric Person
Authentication (pp. 436-446). Springer Berlin Heidelberg.
[54] Breebaart, J., Yang, B., Buhan-Dulman, I., & Busch, C. (2009). Biometric
template protection. Datenschutz und Datensicherheit-DuD, 33(5), 299-304.
[55] Rane, S. (2014). Standardization of Biometric Template Protection.
MultiMedia, IEEE, 21(4), 94-99.
[56] Sun, S. W., Lu, C. S., & Chang, P. C. (2007, January). Biometric template
protection: A key-mixed template approach. In Proceeding IEEE
International Conference Consumer Electronics (pp. 10-14).
[57] Roberge, C. S. D., Stoianov, A., Gilroy, R., & Kumar, B. V. (1999).
Biometric encryption. ICSA Guide to Cryptography, 8.
[58] Sutcu, Y., Li, Q., & Memon, N. (2007, February). How to protect biometric
templates. In Electronic Imaging 2007 (pp. 650514-650514). International
Society for Optics and Photonics.
123
[59] Brindha, V. E. (2012). Biometric Template Security using Dorsal Hand Vein
Fuzzy Vault. J Biomet Biostat, 3(145), 2.
[60] Sutcu, Y., Sencar, H. T., & Memon, N. (2007, April). A geometric
transformation to protect minutiae-based fingerprint templates. In Defense
and Security Symposium (pp. 65390E-65390E). International Society for
Optics and Photonics.
[61] Mirmohamadsadeghi, L., & Drygajlo, A. (2013, September). A template
privacy protection scheme for fingerprint minutiae descriptors. In Biometrics
Special Interest Group (BIOSIG), 2013 International Conference of the (pp.
1-8). IEEE.
[62] Prasad, M. V., & Santhosh Kumar, C. (2014). Fingerprint template protection
using multiline neighboring relation. Expert Systems with Applications,
41(14), 6114-6122.
[63] Jain, A. K., Nandakumar, K., & Nagar, A. (2013). Fingerprint Template
Protection: From Theory to Practice. In Security and Privacy in Biometrics
(pp. 187-214). Springer London.
[64] Teoh, A. B. J., Toh, K. A., & Yip, W. K. (2007). 2^ N discretisation of
biophasor in cancellable biometrics. In Advances in Biometrics (pp. 435-444).
Springer Berlin Heidelberg.
[65] Scheirer, W. J., & Boult, T. E. (2007, September). Cracking fuzzy vaults and
biometric encryption. In Biometrics Symposium, 2007 (pp. 1-6). IEEE.
[66] Draper, S. C., Khisti, A., Martinian, E., Vetro, A., & Yedidia, J. S. (2007,
April). Using distributed source coding to secure fingerprint biometrics. In
Acoustics, Speech and Signal Processing, 2007. ICASSP 2007. IEEE
International Conference on (Vol. 2, pp. II-129). IEEE.
[67] Ratha, N. K., Chikkerur, S., Connell, J. H., & Bolle, R. M. (2007).
Generating cancelable fingerprint templates. Pattern Analysis and Machine
Intelligence, IEEE Transactions on, 29(4), 561-572.
[68] Jeffers, J., & Arakala, A. (2007). Fingerprint alignment for a minutiae-based
fuzzy vault. In Proc Biometric Symposium, BCC, Baltimore, MD.
[69] Farooq, F., Bolle, R. M., Jea, T. Y., & Ratha, N. (2007, June). Anonymous
and revocable fingerprint recognition. In Computer Vision and Pattern
Recognition, 2007. CVPR'07. IEEE Conference on (pp. 1-7). IEEE.
[70] Sutcu, Y., Rane, S., Yedidia, J. S., Draper, S. C., & Vetro, A. (2008, July).
Feature extraction for a Slepian-Wolf biometric system using LDPC codes. In
Information Theory, 2008. ISIT 2008. IEEE International Symposium on (pp.
2297-2301). IEEE.
124
[71] Rathor, K. Iris Collarette Boundary Localization Using 2-D DFT For Iris
Based Biometric System.
[72] Imtiaz, H., & Fattah, S. A. (2011). A spectral domain dominant feature
extraction algorithm for palm-print recognition. International Journal of
Image Processing, 5, 130-144.
[73] Amornraksa, T., & Tachaphetpiboon, S. (2006). Fingerprint recognition using
DCT features. Electronics Letters, 42(9), 522-523.
[74] Imtiaz, H., & Fattah, S. A. (2010, October). A DCT-based feature extraction
algorithm for palm-print recognition. In Communication Control and
Computing Technologies (ICCCCT), 2010 IEEE International Conference on
(pp. 657-660). IEEE.
[75] Badrinath, G. S., Tiwari, K., & Gupta, P. (2012). An efficient palmprint
based recognition system using 1D-DCT features. In Intelligent Computing
Technology (pp. 594-601). Springer Berlin Heidelberg.
[76] Singh, S., Ramalho, M., Correia, P. L., & Soares, L. D. (2012, August). PP-
RIDER: A Rotation-Invariant degraded partial palmprint recognition
technique. In Signal Processing Conference (EUSIPCO), 2012 Proceedings
of the 20th European (pp. 1499-1503). IEEE.
[77] Prungsinchai, S., Khelifi, F., & Bouridane, A. (2013, September). Fourier-
Mellin transform for Robust Image Hashing. In Emerging Security
Technologies (EST), 2013 Fourth International Conference on (pp. 58-61).
IEEE.
[78] Ekinci, M., & Aykut, M. (2007). Palmprint recognition by applying wavelet
subband representation and kernel PCA. In Machine learning and data
mining in pattern recognition (pp. 628-642). Springer Berlin Heidelberg.
[79] Yang, F., Ma, B., Xia Wang, Q., & Yao, D. (2007, June). Information fusion
of biometrics based-on fingerprint, hand-geometry and palm-print. In
Automatic Identification Advanced Technologies, 2007 IEEE Workshop on
(pp. 247-252). IEEE.
[80] Prungsinchai, S. (2014). Robust and secure perceptual hashing in the
transform domain (Doctoral dissertation, Northumbria University).
[81] Derrode, S., & Ghorbel, F. (2001). Robust and efficient FourierโMellin
transform approximations for gray-level image reconstruction and complete
invariant description. Computer Vision and Image Understanding, 83(1), 57-
78.
[82] Khayam, S. A. (2003). The discrete cosine transform (DCT): theory and
Application. Michigan State University.
125
[83] Mallat, S. G. (1989). A theory for multiresolution signal decomposition: the
wavelet representation. Pattern Analysis and Machine Intelligence, IEEE
Transactions on, 11(7), 674-693.
[84] Smith, M. J., & Barnwell III, T. P. (1987). A new filter bank theory for time-
frequency representation. Acoustics, Speech and Signal Processing, IEEE
Transactions on, 35(3), 314-327.
[85] Bhattacharjee, S. K., & Vandergheynst, P. (1999, October). End-stopped
wavelets for detecting low-level features. In SPIE's International Symposium
on Optical Science, Engineering, and Instrumentation (pp. 732-741).
International Society for Optics and Photonics.
[86] Antoine, J. P., & Murenzi, R. (1996). Two-dimensional directional wavelets
and the scale-angle representation. Signal processing, 52(3), 259-281.
[87] Geisler, W. S., Gilbert, J. E., & Ghosh, J. (2005). Perceptually based methods
for robust image hashing.
[88] http://bias.csr.unibo.it/fvc2004/download.asp
[89] Beaudet, P. R. (1978, November). Rotationally invariant image operators. In
International Joint Conference on Pattern Recognition (Vol. 579, p. 583).
[90] Simard, P. Y., Bottou, L., Haffner, P., & LeCun, Y. (1999). Boxlets: a fast
convolution algorithm for signal processing and neural networks. Advances in
Neural Information Processing Systems, 571-577.
[91] Harris, C., & Stephens, M. (1988, August). A combined corner and edge
detector. In Alvey vision conference (Vol. 15, p. 50).
[92] Schneider, M., & Chang, S. F. (1996, September). A robust content based
digital signature for image authentication. In Image Processing, 1996.
Proceedings., International Conference on (Vol. 3, pp. 227-230). IEEE.
[93] Lin, C. Y., & Chang, S. F. (2001). A robust image authentication method
distinguishing JPEG compression from malicious manipulation. Circuits and
Systems for Video Technology, IEEE Transactions on, 11(2), 153-168.
126
[94] Lu, C. S., & Liao, H. Y. (2003). Structural digital signature for image
authentication: an incidental distortion resistant scheme. Multimedia, IEEE
Transactions on, 5(2), 161-173.
[95] Fridrich, J., & Goljan, M. (2000). Robust hash functions for digital
watermarking. In Information Technology: Coding and Computing, 2000.
Proceedings. International Conference on (pp. 178-183). IEEE.
[96] Bhattacharjee, S., & Kutter, M. (1998, October). Compression tolerant image
authentication. In Image Processing, 1998. ICIP 98. Proceedings. 1998
International Conference on (Vol. 1, pp. 435-439). IEEE.
[97] Lin, D. W., & Yang, S. H. (2007). Wavelet-based salient region extraction. In
Advances in Multimedia Information ProcessingโPCM 2007 (pp. 389-392).
Springer Berlin Heidelberg
[98] Lim, J., Kim, Y., & Paik, J. (2009). Comparative analysis of wavelet-based
scale-invariant feature extraction using different wavelet bases. In Signal
Processing, Image Processing and Pattern Recognition (pp. 297-303).
Springer Berlin Heidelberg.
[99] Sebe, N. & Lew, M. S. (2003). Comparing salient point detectors. Pattern
recognition letters, 24(1), 89-96.
[100] Loupias, E., Sebe, N., Bres, S., & Jolion, J. M. (2000, September). Wavelet-
based salient points for image retrieval. In Image Processing, 2000.
Proceedings. 2000 International Conference on (Vol. 2, pp. 518-521). IEEE.
[101] Omidyeganeh, M., Shirmohammadi, S., Laganiere, R., Youmaran, R., &
Javadtalab, A. (2013, July). Face identification using wavelet transform of
SIFT features. In Multimedia and Expo Workshops (ICMEW), 2013 IEEE
International Conference on (pp. 1-6). IEEE.
[102] Kumar, N.A.M., Sathidevi, P.S.,(2012).Image match using wavelet โ Colour
SIFT features. Industrial and Information Systems (pp.1-6).IEEE.
[103] Wang, K., Ren, Z., & Xiong, X. (2008, August). Combination of Wavelet
and SIFT Features for Image Classification Using Trained Gaussion Mixture
Model. In Intelligent Information Hiding and Multimedia Signal Processing,
2008. IIHMSP'08 International Conference on (pp. 79-82). IEEE.
[104] Halawani, A., & Burkhardt, H. (2004, August). Image retrieval by local
evaluation of nonlinear kernel functions around salient points. In Pattern
Recognition, 2004. ICPR 2004. Proceedings of the 17th International
Conference on (Vol. 2, pp. 955-960). IEEE.
127
[105] Tu, Z., & Zhu, S. C. (2002). Image segmentation by data-driven Markov
chain Monte Carlo. Pattern Analysis and Machine Intelligence, IEEE
Transactions on, 24(5), 657-673. Tu, Z., & Zhu, S. C. (2002). Image
segmentation by data-driven Markov chain Monte Carlo. Pattern Analysis
and Machine Intelligence, IEEE Transactions on, 24(5), 657-673.
[106] Jian, M., Dong, J., & Jiang, R. (2007, July). Wavelet-based salient regions
and their spatial distribution for image retrieval. In Multimedia and Expo,
2007 IEEE International Conference on (pp. 2194-2197). IEEE.
[107] Imamoglu, N., Lin, W., & Fang, Y. (2013). A saliency detection model using
low-level features based on wavelet transform. Multimedia, IEEE
Transactions on, 15(1), 96-105.
[108] Tsai, Y. H. (2012). Hierarchical salient point selection for image retrieval.
Pattern Recognition Letters, 33(12), 1587-1593.
[109] Gonzalez, R. C., Woods, R. E., & Eddins, S. L. (2004). Digital image
processing using MATLAB. Upper Saddle River, N. J: Pearson Prentice
Hall.
[110] Porwik, P., & Lisowska, A. (2004). The Haar-wavelet transform in digital
image processing: its status and achievements. Machine graphics and vision,
13(1/2), 79-98.
[111] Roy, S., & Sun, Q. (2007, September). Robust hash for detecting and
localizing image tampering. In Image Processing, 2007. ICIP 2007. IEEE
International Conference on (Vol. 6, pp. VI-117). IEEE.
[112] Chum, O., Philbin, J., & Zisserman, A. (2008, September). Near Duplicate
Image Detection: min-Hash and tf-idf Weighting. In BMVC (Vol. 810, pp.
812-815).
[113] Ke, Y., Sukthankar, R., & Huston, L. (2004, October). Efficient near-
duplicate detection and sub-image retrieval. In ACM Multimedia (Vol. 4, No.
1, p. 5).
[114] Belongie, S., Malik, J., & Puzicha, J. (2002). Shape matching and object
recognition using shape contexts. Pattern Analysis and Machine Intelligence,
IEEE Transactions on, 24(4), 509-522.
[115] http://bias.csr.unibo.it/fvc2002/databases.asp
[116] Lei, B., Tan, E. L., Chen, S., Ni, D., Wang, T., & Lei, H. (2014). Reversible
watermarking scheme for medical image based on differential evolution.
Expert Systems with Applications, 41(7), 3178-3188.
128
[117] Aslantas, V. (2009). An optimal robust digital image watermarking based on
SVD using differential evolution algorithm. Optics Communications, 282(5),
769-777.
[118] Bhatnagar, G., & Raman, B. (2009). A new robust reference watermarking
scheme based on DWT-SVD. Computer Standards & Interfaces, 31(5), 1002-
1013.
[119] Ghouti, L., Bouridane, A., Ibrahim, M. K., & Boussakta, S. (2006). Digital
image watermarking using balanced multiwavelets. Signal Processing, IEEE
Transactions on, 54(4), 1519-1536.
[120] Ramakrishnan, S., Gopalakrishnan, T., & Balasamy, K. (2011). SVD Based
Robust Digital Watermarking For Still Images Using Wavelet Transform.
CCSEA, 155-167.
[121] Hore, A., & Ziou, D. (2010, August). Image quality metrics: PSNR vs. SSIM.
In Pattern Recognition (ICPR), 2010 20th International Conference on (pp.
2366-2369). IEEE.
[122] Tang, Z., Wang, S., Zhang, X., & Wei, W. (2009, June). Perceptual similarity
metric resilient to rotation for application in robust image hashing. In
Multimedia and Ubiquitous Engineering, 2009. MUE'09. Third International
Conference on (pp. 183-188). IEEE.
[123] Wang, Z., Bovik, A. C., Sheikh, H. R., & Simoncelli, E. P. (2004). Image
quality assessment: from error visibility to structural similarity. Image
Processing, IEEE Transactions on, 13(4), 600-612.
[124] Al-Najjar, Y. A. (2012). Der Chen Soong," Comparison of Image Quality
Assessment: PSNR, HVS, SSIM, UIQI. International Journal of Scientific &
Engineering Research, 3(8), 2229-5518.
[125] Hofbauer, H., Rathgeb, C., Uhl, A., & Wild, P. (2012, September). Image
metric-based biometric comparators: A supplement to feature vector-based
Hamming distance?. In Biometrics Special Interest Group (BIOSIG), 2012
BIOSIG-Proceedings of the International Conference of the (pp. 1-5). IEEE.
[126] Keimel, C., & Diepold, K. (2010). Improving the prediction accuracy of
PSNR by simple temporal pooling. In Fifth International Workshop on Video
Processing and Quality Metrics for Consumer Electronics-VPQM (Vol.
2009).
[127] Run, R. S., Horng, S. J., Lai, J. L., Kao, T. W., & Chen, R. J. (2012). An
improved SVD-based watermarking technique for copyright protection.
Expert Systems with Applications, 39(1), 673-689.
129
[128] Braeckman, G., Barri, A., Fodor, G., Dooms, A., Barbarien, J., Schelkens, P.,
& Weng, L. (2012, January). Reduced reference quality assessment based on
watermarking and perceptual hashing. In Sixth International Workshop on
Video Processing and Quality Metrics for Consumer Electronics, Scottsdale,
Arizona (USA).
[129] Gao, X., Lu, W., Tao, D., & Li, X. (2009). Image quality assessment based
on multiscale geometric analysis. Image Processing, IEEE Transactions on,
18(7), 1409-1423.
[130] Li, J., Wu, K., Zhang, X., & Ding, M. (2012). Image quality assessment
based on multi-channel regional mutual information. AEU-International
Journal of Electronics and Communications, 66(9), 784-787.
[131] Weng, L., Braeckman, G., Dooms, A., Preneel, B., & Schelkens, P. (2012,
July). Robust Image Content Authentication with Tamper Location. In
Multimedia and Expo (ICME), 2012 IEEE International Conference on (pp.
380-385). IEEE.
[132] Qin, C., Chang, C. C., & Tsou, P. L. (2012). Perceptual Image Hashing Based
on the Error Diffusion Halftone Mechanism. International Journal of
Innovative Computing, Information and Control, 8(9), 6161-6172.
[133] Wang, Z., & Bovik, A. C. (2009). Mean squared error: love it or leave it? A
new look at signal fidelity measures. Signal Processing Magazine, IEEE,
26(1), 98-117.